# Included academic studies: AI in workshop and meeting facilitation Search date: 2026-09-26. Protocol: `../LITERATURE-SEARCH-PROTOCOL.md` (sources 1, 2, 3 and 5; grey literature is reported separately). Screening log: `search-log.csv` (one row per screened record; `record_key` links rows to the entries below). ## How the search was run **OpenAlex** (api.openalex.org, `from_publication_date:2022-01-01`). The literal protocol queries were run on title and abstract to record totals, but four of them return tens of thousands of hits (the terms "facilitation", "meeting", "workshop", "agenda" and "retrospective" are generic), and OpenAlex ranks filtered results by citations, not relevance. The protocol queries were therefore split into narrower Boolean sub-queries whose complete result sets were screened (every record, not only the first 50): | Query | Field | Boolean string (abridged) | Hits (all screened) | |---|---|---|---| | OA-Q1a | title | (facilitation OR facilitator OR facilitators OR facilitating) AND (LLM terms incl. "AI assistant") | 348 | | OA-Q1b | title | (meeting OR meetings) AND (LLM terms) | 107 | | OA-Q1c | title | (workshop OR workshops) AND ("large language model" OR LLM OR "generative AI" OR ChatGPT) | 265 | | OA-Q1c2 | title | (workshop OR workshops) AND "AI assistant" | 2 | | OA-Q1d | title+abstract | (facilitator OR facilitators) AND (LLM terms) AND (workshop OR meeting OR group), ranked by relevance to "facilitator"; top 100 screened of 2,335 | 100 of 2,335 | | OA-Q2 | title+abstract | ("group ideation" OR brainstorming OR "collective creativity") AND ("large language model" OR "generative AI") [protocol query as written] | 741 | | OA-Q3 | title+abstract | (deliberation OR deliberative) AND ("large language model" OR "AI facilitator" OR "AI moderator") [protocol query as written] | 963 | | OA-Q4 | title+abstract | ("meeting planning" OR "workshop design" OR "meeting agenda" OR "workshop planning") AND (AI OR LLM OR "large language model" OR "generative AI") | 117 | | OA-Q5 | title+abstract | ("design thinking" OR "design sprint" OR "sprint retrospective" OR "agile retrospective") AND ("generative AI" OR LLM OR "large language model") | 390 | Totals for the literal protocol queries (title and abstract, 2022+): Q1 31,170; Q2 741; Q3 963; Q4 10,575; Q5 2,681. Records retrieved and screened: 3,033 (2,942 unique after removing records returned by more than one sub-query). **Semantic Scholar** (graph API). The unauthenticated relevance endpoint was rate-limited for most of the day (HTTP 429 on almost every call); only four of the fifteen planned relevance queries completed (S2-S1a "facilitator large language model workshop", S1b "generative AI meeting facilitation", S1c "ChatGPT workshop facilitation", S1d "AI assistant meetings large language model"; first 50 each, `year=2022-2026`). The remaining coverage came from the bulk search endpoint, which accepts Boolean syntax and reports totals. Totals for the protocol queries in bulk form: Q1 17,671; Q2 350; Q3 461; Q4 4,367; Q5 1,466. Complete result sets screened: S2-B2 (protocol Q2, 350), S2-B3 (protocol Q3, 461), S2-B1b (meeting terms + LLM, 64), S2-B1c (group discussion/decision + LLM + facilitat*/moderat*/mediat*, 132), S2-B1e (facilitator + workshop/meeting + LLM, 303), S2-B4n (meeting planning/workshop design/agenda design + AI, 55), S2-B5n (design sprint/retrospective/innovation workshop + LLM, 19). Records retrieved: 1,584 (1,525 unique within Semantic Scholar). **Google Scholar cross-check.** Google Scholar has no API; the key queries were run through web search instead (about fifteen searches, covering LLM facilitators in group discussion, facilitators' use of generative AI, LLM group brainstorming, AI moderators in deliberation, meeting agenda planning, sprint retrospectives, design sprints, citizens' assemblies, German-language "KI Moderation Workshop Studie", Copilot in meetings, AI teammates in hidden-profile tasks, LLM focus-group moderators, design-thinking workshops, facilitator interview studies and AI note-takers). This is not the "first 50 results per query" the protocol specifies; it is a cross-check. Fifteen items the two APIs had not surfaced were screened (keys GS1–GS15). **Citation check.** Round 1: references and citing works (OpenAlex `referenced_works` and `cites:`) of the 62 records provisionally included from OpenAlex: 826 new records from 2022 or later. Round 2: the same for the 51 includes found in round 1 and in the Google Scholar cross-check: 1,727 new records. Studies first found in round 2 or in Semantic Scholar were not themselves citation-checked (a third round was not run). **Screening rules applied** (from the protocol, made operational): generative AI/LLM involved; empirical with stated method and sample; at least two humans interacting as a group with the AI, *or* facilitators/meeting professionals/organisers reporting their own AI use or using AI to plan sessions, *or* an offline evaluation of an LLM performing a facilitation function on group-discussion data; systematic or scoping reviews allowed and labelled; classroom use excluded unless the AI takes part in a structured group session with a group-level outcome. Doshi & Hauser (2024) is retained as the protocol requires. Records whose full text was not accessible and whose abstract gives no sample size (or leaves the group format or the use of generative AI unclear) are listed as **not assessable** rather than counted as included. Exclusion codes are defined in the PRISMA section. **How to read the entries.** Each entry was written from the full text where one was openly available, otherwise from the abstract ("Full text read" says which). Page numbers refer to the version read, which is sometimes an arXiv or repository copy with different pagination from the published version; the entry says so. Numbers are as printed in the source. Where a source was internally inconsistent, the entry notes it. Nothing below has been checked against supplementary materials unless stated. **Dependencies to keep in mind when counting evidence.** (1) Eleven included papers on Conversational Swarm Intelligence / Thinkscape / Hyperchat AI (OA2022, OA2071, OA2121, OA2213, OA2540, OA1308, C23, C84, C85, C672, R86) come from one company (Unanimous AI) and cite each other's results; count them as one programme. (2) C18 and S2-567 both study the "Koala" agent in group brainstorming; OA735 and S2-569 both study the "Collaborative Canvas" and may share participants. (3) C82 (PNAS) is the published version of the preprint in OA2154, which is logged as a duplicate. (4) GS15 and OA1567 (logged as not assessable) are by the same team and may share data. (5) OA1297 and OA1461 report the same classroom brainstorming programme from different angles. (6) R964 and R1506 share five authors. (7) C5 (Habermas Machine) builds on C697 by the same group. (8) OA1450 and OA1723 come from the same lab and paradigm. ## 1. AI facilitating, moderating or mediating live group sessions #### OA136 — LLM-facilitated group decision making (hidden profile experiment) - **Citation:** Alsobay, M., Rothschild, D. M., Hofman, J. M., & Goldstein, D. G. (2026). Bringing Everyone to the Table: An Experimental Study of LLM-Facilitated Group Decision Making. Proceedings of the ACM on Human-Computer Interaction, 10(6), Article CSCW059, 1–36. https://doi.org/10.1145/3816907 - **Type:** journal (PACM HCI, CSCW) — peer-reviewed: yes - **Full text read:** yes (arXiv:2508.08242v2, 2 Jul 2026, formatted as the PACM HCI article; https://arxiv.org/pdf/2508.08242). Page numbers = article pages CSCW059:N. - **Method:** Pre-registered (AsPredicted #192061) randomized experiment, four between-group conditions: no facilitation, one-time message prompting information sharing, human facilitator, LLM facilitator; confirmatory one-sided t-tests with Benjamini-Hochberg adjustment (§3, pp. 5, 10). - **n:** 1,475 Prolific participants in 281 live groups of five decision makers (a sixth participant acted as human facilitator in that condition); ~70 valid groups per condition (pp. 1, 5, 11, 24). - **Setting:** Online text chat; hidden profile task (choose the optimal city for a hypothetical sporting event); LLM facilitator = GPT-4o (abstract, §3). - **Findings:** 1. LLM facilitation increased the number of facts shared vs. no facilitation (Δ = 2.9, t(138) = 3.58, p < 0.001, B-H adj. p = 0.001), vs. one-time message (Δ = 1.58, p = 0.021, adj. p = 0.035) and vs. human facilitation (Δ = 2.47, t(138) = 3.38, p < 0.001, adj. p = 0.001); neither the message (Δ = 1.32, adj. p = 0.082) nor the human facilitator (Δ = 0.43, p = 0.31) significantly increased sharing vs. none (§4.1, p. 11). The gain came from raising the minimum individual contribution (Δ = 0.41 facts vs none, 95% CI [0.145, 0.684], p = 0.003) (p. 12). 2. No effect on decision quality: the optimal option was chosen by 31% (none), 21% (message), 30% (human) and 23% (LLM) of groups, χ²(3, N = 281) = 2.85, p = 0.42 (§4.3, p. 13). 3. Groups rated the LLM facilitator higher than the human facilitator on eliciting information (Δ = 0.41, 95% CI [0.22, 0.60], p < 0.001) and summarising (Δ = 0.59, 95% CI [0.39, 0.79], p < 0.001), but also as more distracting (Δ = 0.38, 95% CI [0.21, 0.56], p < 0.001), citing repetitive interventions and disrupted flow (p. 16). - **Relevance:** (a) — | (b) LLM vs. human facilitator running live small-group decision discussions | (c) more information shared but no improvement in decision quality - **Caveats:** Human facilitators were Prolific crowdworkers, not professional facilitators; short online text task; hidden-profile paradigm; higher exclusion rate in the human condition (non-significant, χ²(3, N = 317) = 4.56, p = 0.21) (p. 11). - **Screening check:** Meets criteria (pre-registered experiment, 5-person groups interacting with an LLM facilitator, method and n reported). #### OA290 — LLM facilitation in real-time group deliberation (charity allocation) - **Citation:** Parisi, A., Thain, N., Hallak, A., Tsai, V., & Qian, C. (2026). Real-Time Group Dynamics with LLM Facilitation: Evidence from a Charity Allocation Task. In The 2026 ACM Conference on Fairness, Accountability, and Transparency (FAccT '26). https://doi.org/10.1145/3805689.3812253 - **Type:** conference — peer-reviewed: yes - **Full text read:** yes (arXiv:2605.14097v2, 29 May 2026, https://arxiv.org/pdf/2605.14097; page numbers = preprint pages) - **Method:** Two incentive-compatible online experiments with real donations ($7,200); within-group treatments in randomised order. Study 1 compares three facilitator models under a minimal prompt; Study 2 compares two facilitator strategies (Principles-based, Summarization-based; Gemini 2.5 Flash) against a no-facilitator baseline; power analyses for consensus change (§3). - **n:** N = 879 in total: Study 1 204 participants in 68 groups (258 recruited); Study 2 675 participants in 225 groups (796 recruited); groups of three; 846 unique participants (3% took part twice) (abstract; §3 p. 7; App. E p. 23). - **Setting:** Prolific participants, real-time text deliberation (five-minute discussion rounds) allocating a donation budget across charities; Gemini 2.5 Flash, Claude 4.5 Haiku, GPT-5 mini (§3). - **Findings:** 1. No statistically significant aggregate effect of LLM facilitation on group consensus in either study (effects consistently positive; possible ceiling, round-one median consensus 91.1, IQR [68.3, 100]) (§4.1, p. 8). 2. Algorithmic steering: Principles and Summarization facilitators lowered allocations to the Global Fund for Women by −5.5 and −3.8 percentage points vs. baseline (p < 0.001, Welch's t-test); mechanism: facilitators restated and formalised outlier proposals (§4.1, pp. 8–9). 3. Participants significantly preferred LLM facilitation over no facilitator in Study 2 (χ², p < 0.001) and preferred Gemini and GPT over Haiku in Study 1 (p < 0.05), often citing inclusivity (coded theme counts N = 90 among Summarization-preferred and N = 106 among Principles-preferred responses), yet neither survey nor transcript measures of participation equity improved; LLM facilitation raised total message volume through LLM messages, while humans contributed fewer turns (~10–14 vs 15.0 per round) ("illusion of inclusion"; §4.2–4.3, pp. 9–11, 13). - **Relevance:** (a) — | (b) LLM facilitating live small-group deliberations; model and strategy comparisons | (c) no consensus gain, measurable steering of the group decision, perceived but not actual inclusion - **Caveats:** Very short discussions (5 min), three-person crowdworker groups; steering and dynamics analyses exploratory (not separately powered); no human-facilitator arm. - **Screening check:** Meets criteria (randomised experiments, 3-person groups interacting with LLM facilitators, method and n reported). #### C5 — AI can help humans find common ground in democratic deliberation (Habermas Machine) - **Citation:** Tessler, M. H., Bakker, M. A., Jarrett, D., Sheahan, H., Chadwick, M. J., Koster, R., Evans, G., Campbell-Gillingham, L., Collins, T., Parkes, D. C., Botvinick, M., & Summerfield, C. (2024). AI can help humans find common ground in democratic deliberation. Science, 386(6719), eadq2852. https://doi.org/10.1126/science.adq2852 - **Type:** journal — peer-reviewed: yes - **Full text read:** yes (publisher PDF of the article, obtained from a third-party copy at https://www.rivista.ai/wp-content/uploads/2024/10/science.adq2852.pdf; supplementary materials not read). Page numbers below are the article's "x of 9" pages. - **Method:** Series of structured online experiments: small groups (typically 5–6) privately write opinions, an LLM system generates candidate group statements, participants rank them, write critiques, and rank revised statements (caucus mediation); comparisons against trained, incentivized human mediators, an opinion-exposure (no mediation) control cohort, ablation cohorts, and a virtual citizens' assembly; mixed-effects models (Fig. 1; Results RQ1–RQ4, p. 1–5 of 9). - **n:** Total N = 5734 UK adults (Summary; p. 1 of 9); human-mediator comparison: 75 six-person groups (n = 439, one member per group as mediator) (p. 2 of 9); main task cohorts 1–3 combined: n = 1692 in 349 groups, 1047 deliberation rounds (p. 3 of 9); opinion-exposure control n = 372 in 75 groups (p. 4 of 9); virtual citizens' assembly n = 200 demographically representative participants (Sortition Foundation), three weekly 1-h sessions (p. 5 of 9). - **Setting:** UK residents (mostly crowdsourced) deliberating on divisive UK policy questions (e.g., Brexit, immigration, voting age); "Habermas Machine" = fine-tuned Chinchilla LLM generator + personalized reward model + social-choice aggregation (p. 1 of 9). - **Findings:** 1. AI statements were preferred over human mediators' statements 56% vs 44% of the time (predicted probability 0.575, 95% CI 0.535–0.614, P < 0.001) and rated higher in quality (OR = 1.54, P < 0.001) and endorsement (OR = 1.36, P < 0.001) (RQ1, p. 2 of 9). 2. Group agreement increased by about eight percentage points after AI-mediated deliberation (e.g., cohort 2: b = 0.093, P < 0.001); unanimous groups rose from 22.8% to 38.6% in cohort 1; no change with mere opinion exposure (b = −0.005, P = 0.64), significant interaction (P = 0.0058) (RQ2, p. 4 of 9). 3. Statements incorporated minority critiques: initial majority grew in 29–32% of rounds and minority in 20–26%; results (endorsement, increased agreement) replicated in the representative virtual citizens' assembly (RQ3, p. 4 of 9; RQ4, p. 5 of 9). - **Relevance:** (a) — | (b) LLM performs a core mediator/facilitator function (drafting and revising group consensus statements), outperforming human mediators | (c) AI mediation reduced within-group division and raised agreement/consensus - **Caveats:** Text-based, asynchronous-style caucus mediation, not live facilitated discussion; human mediators had no critique round; mostly crowdsourced UK sample; Google DeepMind authors; supplementary analyses not verified. - **Screening check:** Meets criteria — empirical, groups of humans deliberating with an LLM mediator, method and n reported. #### OA284 — PTFA: Parallel Thinking-based Facilitation Agent (Six Thinking Hats) - **Citation:** Gu, W., Li, Z., Buermann, J., Dilkes, J., Michailidis, D., Hasegawa, S., Yazdanpanah, V., & Stein, S. (2026). PTFA: An LLM-Based Agent that Facilitates Online Consensus Building Through Parallel Thinking. Lecture Notes in Computer Science, pp. 125–141. https://doi.org/10.1007/978-981-95-7078-2_9 - **Type:** conference (LNCS chapter) — peer-reviewed: yes - **Full text read:** yes (arXiv:2503.12499v2, 22 Oct 2025, https://arxiv.org/pdf/2503.12499; page numbers = preprint pages) - **Method:** Pilot user study, within-group: each group held two 20-min online text discussions, one with a fixed-schedule baseline facilitator (three timed messages) and one with PTFA (LLM performing all six Six Thinking Hats roles, intervention decisions every 30 s); post-discussion 7-point surveys plus qualitative transcript analysis; no inferential statistics reported (§4.1–4.2). - **n:** 48 participants in 16 groups of 3 (recruited at University of Southampton); 32 discussions, 1,669 posts / 16,656 words, of which the facilitator contributed 217 posts / 3,459 words (§4.1–4.2, pp. 7–8). - **Setting:** University tutorial rooms, custom online text platform; topics "decide one activity to do together" and "decide one film to watch together"; OpenAI ChatGPT Assistant API (model version not verified) (§3, §4.1). - **Findings:** 1. Facilitator helped consensus decision-making: agreement "slightly above 50%" for PTFA vs "slightly below 50%" for the baseline; authors characterise ratings of both as mixed (§4.2, p. 10; exact percentages only in Fig. 4b, not verified). 2. User experience and agreement with the reached consensus were positive in both conditions but "slightly dampened" with PTFA, which authors attribute to problematic LLM patterns such as difficulty keeping to phases (p. 10). 3. Qualitative: the agent contributed perspectives participants neglected (emotional reactions, strengths/weaknesses, idea generation), but kept asking for new ideas until the end, sometimes ignored direct questions, and sometimes responded after every comment; in one discussion late new ideas split preferences (pp. 10–13). - **Relevance:** (a) — | (b) LLM agent actively facilitating small-group online consensus building with a structured technique | (c) mixed effects on perceived consensus support; LLM adds ideas/perspectives but can disrupt convergence - **Caveats:** Pilot; trivial low-stakes tasks; baseline is a 3-message timer, not a human facilitator; results descriptive; student/university sample. - **Screening check:** Meets criteria (empirical, 3-person groups interacting with an LLM facilitator, method and n reported). #### OA148 — When to facilitate: human vs. LLM facilitator tendencies - **Citation:** Tsirmpas, D., Korre, K., & Pavlopoulos, J. (2026). To Facilitate or not to Facilitate: Human and LLM Facilitator Tendencies in Online Discussions. arXiv:2607.28643 (v1, 27 May 2026). https://doi.org/10.48550/arxiv.2607.28643 - **Type:** preprint — peer-reviewed: no (preprint) - **Full text read:** yes (https://arxiv.org/pdf/2607.28643, v1) - **Method:** Offline evaluation. (1) Survey in which human experts and several open-source LLMs (LLM-as-judge) judge, for 6-comment discussion chunks, whether positive/negative reinforcement (facilitative intervention) is needed; (2) intervention-prediction and facilitation-detection classification tasks on a new aggregated corpus (PEFK), comparing six open-source LLMs with a fine-tuned ModernBERT classifier (§3–5). - **n:** 11 human experts recruited (computational social science/NLP background; 4 male, 7 female), 1 did not complete, so 10 analysed; 1,224 discussion chunks sampled, 721 retained after malformed excerpts flagged (§4.1, p. 4); 2,000 test instances for LLM prediction task (§5.1, p. 6). - **Setting:** Existing multi-party discussion datasets (online forums, deliberation, oral debates such as IQ2, Fora, CeRI, WikiTactics etc.) standardised into PEFK; LLMs include Llama 8B/70B and Qwen 32B among others (App. C.1; full list not verified). - **Findings:** 1. Humans are cautious facilitators: most participants chose not to intervene in most cases; positive-intervention rates cluster at ~20–30% and 35–40% (§4.2, p. 5, Finding 1). LLMs are "more intrusive and more consistent", treating neutral discussions as opportunities for positive reinforcement (Finding 2, p. 6). 2. There were 36 discussion instances where no human agreed with the LLM majority vote (LLM agreement ≥ 0.66) (§4.3, p. 6). 3. Deciding when not to facilitate is easy for both humans and LLMs (high F1 for the negative class) but deciding when to facilitate is not (low positive-class F1); out-of-the-box LLMs with generic prompting are less consistent than encoder-only classifiers (Findings 3–4, p. 7). Example: Qwen32B F1p = 0.904 on CeRI but Llama70B F1p = 0.500 on WT in facilitation detection (§5.3, p. 7). - **Relevance:** (a) — | (b) whether LLMs can judge when a facilitator should intervene in group discussions (over-intervention tendency) | (c) — - **Caveats:** Human raters are NLP/CSS experts given facilitation training, not practising facilitators (the abstract calls them "expert facilitative participants"); offline judgments, no live groups; mostly online text discussions; preprint. - **Screening check:** Meets criteria as an offline evaluation of LLMs performing a facilitation function (intervention timing) on group-discussion data, with method and n reported. #### OA77 — LLMs as automated facilitators for consensus (Divergence to Alignment) - **Citation:** Triantafyllopoulos, L., & Kalles, D. (2025). From Divergence to Alignment: Evaluating the Role of Large Language Models in Facilitating Agreement Through Adaptive Strategies. Future Internet, 17(9), 407. https://doi.org/10.3390/fi17090407 - **Type:** journal — peer-reviewed: yes - **Full text read:** yes, but the preprint version only (arXiv:2503.15521v1, "From Divergence to Consensus: ...", https://arxiv.org/pdf/2503.15521). The MDPI PDF/HTML returned HTTP 403. The published abstract (via OpenAlex) adds "A one-way ANOVA confirmed that differences in performance between models were statistically significant", which does not appear in the preprint; the ANOVA statistics are not verified. Page numbers below are PDF pages of the arXiv preprint. - **Method:** Field experiment with a custom multi-user chat system; LLM generates consensus proposals, participants accept/reject and give feedback, LLM iterates using adaptive strategies (clarify, summarise, propose compromise); outcome = cosine similarity between participants' initial positions and accepted consensus proposal, and number of iterations (§3–5). - **n:** 30 Hellenic Open University Informatics students volunteered; 75 sessions conducted (24 Aug–30 Sep 2024), evenly distributed across three LLMs; 124 calculated similarity values (ChatGPT 4.0: 50, Mistral Large 2: 40, AI21 Jamba: 34). Agreement was reached "between only two participants" per discussion (pp. 16–17, 22). - **Setting:** Pairs of students discussing six controversial SDG-related questions (health, education, climate, water) in Greek; LLM facilitators ChatGPT 4.0, Mistral Large 2, AI21 Jamba-Instruct (§4, Table 2). - **Findings:** 1. Average cosine similarity between users' initial positions and the final accepted proposal: ChatGPT 4.0 0.701, AI21 Jamba 0.6126, Mistral Large 2 0.5807; overall 0.6381 (Table 3, p. 17). 2. ChatGPT 4.0 reached consensus in the first round (without feedback) in 50 cases, vs. below 40 for the other two; with ChatGPT no discussion needed more than three iterations, whereas Mistral and Jamba sometimes needed four to five (§5.2, pp. 19–20). 3. Authors' limitations: cosine similarity may not capture qualitative agreement; only two participants per agreement, all students; discussions in Greek (§6, p. 22). - **Relevance:** (a) — | (b) LLM as real-time consensus facilitator in multi-user chat; compares models | (c) consensus proposals and speed of agreement in dyadic group decisions - **Caveats:** Groups of two; no human-facilitator or no-facilitator control; outcome is embedding similarity, not decision quality; convenience student sample; published version may contain additional statistics (ANOVA) not checked here. - **Screening check:** Meets criteria (empirical, 2 humans interacting with an LLM facilitator, method and n reported). #### OA742 — EchoMind: LLM real-time issue mapping for group discussions - **Citation:** Chen, W., Yu, C., Wang, Y., Chen, M., Xu, Y., & Shi, Y. (2025). EchoMind: Supporting Real-time Complex Problem Discussions through Human-AI Collaborative Facilitation. Proceedings of the ACM on Human-Computer Interaction, 9(7), Article CSCW406, 38 pp. https://doi.org/10.1145/3757587 - **Type:** journal (PACM HCI / CSCW) — peer-reviewed: yes - **Full text read:** yes (author copy, https://pi.cs.tsinghua.edu.cn/wp-content/uploads/2025/10/Weihao_CHEN-EchoMind-CSCW.pdf; pages = printed CSCW406:N) - **Method:** Wizard-of-Oz formative study, then within-subjects user study: each team ran two 30-min face-to-face sessions, one with EchoMind (issue map) and one with AutoDoc (LLM appending key points to a document), order counterbalanced; questionnaires (Wilcoxon signed-rank), NASA-TLX, interviews, blind-ish transcript coding into a Discussion Quality Index (§3, §5) - **n:** formative: 7 participants in 2 research teams + 2 wizards (§3, p. 6); main study: 16 participants in 4 teams (8 sessions), one member per team acting as facilitator (§5.2, §5.4, p. 13) - **Setting:** Tsinghua campus recruits (13 undergraduates, 3 graduates), meeting room with 65-inch shared screen, startup-style product design task (smart products for the elderly), Chinese; GPT-4o generates issue/position nodes (§4, p. 11; §5.2–5.4, p. 13) - **Findings:** 1. Discussion Quality Index 55.57 with EchoMind vs. 44 with AutoDoc; participants felt they addressed topics more effectively (p < .05, W = 9.0), were more satisfied with quality (p < .01, W = 0.0), and perceived greater depth (p < .05) and breadth (p < .05) and a clearer sense of direction (p < .05, W = 10.0) (§6.1.1, p. 15). 2. Facilitators showed 10.25 facilitation behaviours per session with EchoMind vs. 8.75 with AutoDoc, concentrated mid-session; fewer manual edits (13.75 ops/146.5 chars vs. 17 ops/624.5 chars); 14 of 16 would use EchoMind again (§6.1.3–6.1.5, pp. 17–18). 3. Limits of AI facilitation: of 24 focus switches only 7 used system-generated issues (17 entered manually); GPT-4o predicting when to switch focus reached 28% precision and 58.33% recall — "a poor match with the actual user intent" (§6.2, p. 19). - **Relevance:** (a) human facilitator working with AI map (human-AI co-facilitation) | (b) LLM structuring live face-to-face discussion | (c) discussion quality, depth/breadth (subjective + coded) - **Caveats:** only 4 teams, student sample, authors described the comparison as exploratory; coders could partly infer condition; single task. - **Screening check:** Meets criteria — empirical within-subjects study, n=16 in 4 teams reported, groups discussing with GPT-4o support, English. #### OA736 — LLM-powered devil's advocate in AI-assisted group decisions - **Citation:** Chiang, C.-W., Lu, Z., Li, Z., & Yin, M. (2024). Enhancing AI-Assisted Group Decision Making through LLM-Powered Devil's Advocate. Proceedings of the 29th International Conference on Intelligent User Interfaces (IUI '24), pp. 103–119. https://doi.org/10.1145/3640543.3645199 - **Type:** conference — peer-reviewed: yes - **Full text read:** yes (author copy, https://mingyin.org/paper/IUI-24/devil.pdf; PDF page p. X = proceedings page 102+X) - **Method:** randomized between-subjects online experiment, 5 treatments (Control; Static-AI; Static-Majority; Dynamic-AI; Dynamic-Majority = interactivity × target of objection), groups of three chat-deliberating on recidivism predictions with an AI recommendation (RiskComp); regression/ANOVA; qualitative coding of chat logs (κ = 0.75) (§3, §4) - **n:** 350 participants (Prolific) in 120 groups (Control 24, Static-AI 28, Static-Majority 23, Dynamic-AI 28, Dynamic-Majority 17); groups started with 3 members, some finished with 2 (§4, p. 8 of PDF = p. 110; fn. 7) - **Setting:** online text-chat group decision making; devil's advocate powered by OpenAI GPT-3.5-turbo (§1 p. 2 = p. 104; §3.2 p. 5 = p. 107) - **Findings:** 1. Groups with the interactive devil's advocate challenging the AI recommendation (Dynamic-AI) had higher standardized decision accuracy than Control (β = 0.135, SE = 0.068, 95% CI [0.002, 0.269], p = 0.047); no devil's-advocate type significantly changed reliance on wrong AI recommendations (§4.1, pp. 8–9 of PDF = pp. 110–111). 2. Interactive devil's advocates were rated more collaborating (F(1,268) = 9.757, p = 0.002, η² = 0.035) and higher quality (F(1,268) = 11.211, p = 0.001, η² = 0.040); workload was similar across treatments except self-perceived performance (F(4,328) = 2.417, p = 0.048), lowest in Dynamic-AI (§4.4, pp. 9–10 of PDF = pp. 111–112). 3. Dynamic-AI groups discussed longer (β = 198.25 words, SE = 55.31, p < 0.001) but rated teammates' information as less precise and less useful than Control (e.g., usefulness M = 3.821 vs. 4.410, p < 0.001); interactive DAs sometimes invited quiet members to speak, but repetitive arguments led some groups to ignore them (§4.4–4.5, pp. 9–12 of PDF = pp. 111–114). - **Relevance:** (a) — | (b) LLM taking a structured facilitation role (devil's advocate) inside group deliberation | (c) direct: group decision accuracy and reliance on AI - **Caveats:** crowdworkers, text chat, recidivism task; accuracy effects small and several only marginal (p ≈ 0.06–0.08). - **Screening check:** Meets criteria — randomized experiment, n=350/120 groups, ≥2 humans deliberating with an LLM agent, English. #### C286 — LLM devil's advocate against majority influence in group decisions - **Citation:** Lee, S., Hwang, S., Kim, D., & Lee, K. (2025). Conversational Agents as Catalysts for Critical Thinking: Challenging Social Influence in Group Decision-making. Extended Abstracts of the CHI Conference on Human Factors in Computing Systems (CHI EA '25), Yokohama, pp. 1–12. https://doi.org/10.1145/3706599.3719792 (arXiv:2503.14263) - **Type:** conference (extended abstract / late-breaking work) — peer-reviewed: yes (lighter review than full papers). The full text read is the arXiv version. - **Full text read:** yes (https://arxiv.org/pdf/2503.14263) - **Method:** Mixed-methods experiment. Communication condition (baseline chat vs. an LLM devil's advocate posting counterarguments after every 8 messages) was within-subjects, and role (3 high-power "seniors" vs. 1 low-power "junior") was between-subjects. Measures were 7-point scales, NASA-TLX, robust regression, and exit interviews (p. 2–4 of PDF, §3). Counterbalancing of condition order: not verified. - **n:** 48 Korean participants (aged 19–39, M = 26.17; 31 F, 17 M) in 12 groups of 4. Each group did two 20-minute tasks, one per condition (p. 3 of PDF, §3.1, §3.3). - **Setting:** Anonymous online chat on a custom platform. Corporate decision tasks (promotion choice, contract partner) with seniors primed toward stability and juniors toward ambition to create majority–minority dissent. The devil's advocate is multi-agent and GPT-4o based (p. 3–4 of PDF). - **Findings:** 1. Satisfaction with the decision-making process rose from 5.10 (SD 1.52) to 5.55 (SD 1.16) (β = 0.570, p < 0.01) and satisfaction with decision outcome quality from 5.31 (SD 1.55) to 5.89 (SD 1.19) (β = 0.869, p < 0.01). Both increases were significant for juniors and seniors (p. 5 of PDF, §4). 2. Psychological safety rose slightly, from 5.38 to 5.65, which was not significant. Juniors remained significantly less psychologically safe than seniors in both conditions (treatment β = −2.37, p < 0.0001) (p. 4–5 of PDF). 3. Cognitive workload was unchanged (NASA-TLX 3.93 to 4.12, p = 0.883). Juniors said the AI reduced their isolation ("It wasn't just me who had a different opinion") and "gave a little more power to minority opinions" (p. 5 of PDF). - **Relevance:** (a) — | (b) LLM agent acting as an in-discussion devil's advocate (a facilitation technique) | (c) higher process and outcome satisfaction with no added workload; power asymmetries persist - **Caveats:** Extended abstract. Small n (12 groups). Artificial role manipulation. Satisfaction is self-reported; objective decision quality was not measured. Korean sample. The same group of authors has a related devil's-advocate paper on design fixation (arXiv 2406.11125). - **Screening check:** Meets criteria (empirical, groups of 4 humans deciding with an LLM agent in the discussion, method and n reported, English). #### S2-243 — LLM-powered dissenting-minority support in power-imbalanced group decisions - **Citation:** Lee, S., Hwang, S., Kim, M., Kim, D., & Lee, K. (2026). Investigating LLM-Powered Dissenting Minority Support in Power-Imbalanced Group Decision-Making: Counterargument and Mediation as Intervention Strategies. Proceedings of the ACM on Human-Computer Interaction, 10(6), Article CSCW106, 40 pages (October 2026). https://doi.org/10.1145/3816954 (also arXiv:2606.31762). The same study is reported in S2-252 and S2-268; count once. - **Type:** journal (PACM HCI, CSCW issue) — peer-reviewed: yes - **Full text read:** yes (arXiv 2606.31762v1 PDF of the PACM HCI version, pdf3/S2-243.pdf; page numbers as printed, CSCW106:N) - **Method:** Mixed-method controlled lab experiment with online text-chat group discussions. Participant type (majority vs minority) is between-subjects. LLM support type is partially within-subjects: each group completed a Baseline block plus either AIGC (AI-generated counterarguments) or AIMM (AI-mediated private messages paraphrased into the chat, blended with AIGC), with condition and task order counterbalanced. Analysis used robust linear mixed models (rlmer) with Bonferroni post-hoc tests, plus semi-structured exit interviews (pp. 3, 12–14; §4.4–4.5, §5). - **n:** 96 Korean participants (age M = 26.60, SD = 5.21; 61 female, 35 male) in 24 groups of four. Each group had 3 randomly assigned high-power "Seniors" (majority) and 1 "Junior" (minority). Each condition included at least 48 participants. Each group ran two 20-min decision tasks; sessions lasted about 90–105 min (p. 11, §4.2; p. 12, §4.4). - **Setting:** Custom online chat platform with role-played corporate decisions (team-leader promotion; contractor selection), each with conservative, challenging and neutral options designed to split the roles. OpenAI GPT-4o in a multi-agent pipeline (Summary Agent, Paraphrase Agent, Socratic "Devil's Advocate" Conversation Agent, duplicate checker). RAG was used only for paraphrasing juniors' direct messages (p. 7–8, §3.2; p. 11–12, §4.3). - **Findings:** 1. AIMM backfired on psychological safety for the minority. Juniors' psychological safety was lowest in AIMM (M = 3.17, SD = 1.53) vs Baseline (4.25) and AIGC (4.08); AIMM vs Baseline β = −1.40, SE = 0.28, p < .001. Marginalization was highest in AIMM (M = 4.42 vs 3.46 Baseline, 2.92 AIGC; β = 0.96, p < .001). Seniors stayed stable. Interviews: seniors ignored the AI-voiced input ("It's an AI, so I just kind of ignored it") (p. 16, §5.2). 2. Juniors typed more in AIMM (708.62 characters vs 577.62 Baseline and 535.42 AIGC; β = 129.95, p < .05, not significant after Bonferroni). Message counts did not differ by condition or role (p. 17, §5.3, Table 2). 3. AIGC improved the minority's experience. Juniors' overall satisfaction with the process was 4.92 in AIGC vs 3.79 in Baseline and 2.92 in AIMM. Outcome satisfaction was 5.08 (AIGC) vs 3.25 (AIMM). Juniors rated the AI's fairness 5.58 in AIGC vs 4.00 in AIMM (significant) (pp. 19–21, §5.4). Final group decisions still matched senior preferences about 80% of the time (p. 14, §5). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: a GPT-4o devil's-advocate/mediator agent intervenes in live group discussions to rebalance participation under hierarchy | (c) AI and group ideation/decisions: AI counterarguments improved minority satisfaction and atmosphere; anonymous AI revoicing raised minority participation but lowered psychological safety and satisfaction; decisions mostly followed the majority (~80%) - **Caveats:** Role-play with induced power (Seniors decided a bonus for the Junior; everyone received equal pay). Text chat only, Korean sample with high AI familiarity (53.1% had used AI in groups). The number of juniors per AI condition is small (not stated explicitly; 24 juniors split across AIGC/AIMM). The same study appears in S2-252 and S2-268 (duplicate reports). - **Screening check:** Meets criteria. The paper is empirical, with groups of 4 humans interacting with a GPT-4o agent in live group decision-making. It reports its method and n (96 participants / 24 groups) and is in English. #### C171 — AI feedback agent ("Emily") for meeting inclusion - **Citation:** Houtti, M., Zhou, M., Runningen, D., Sunil, S., Porat, L., Kaur, H., Terveen, L., & Chancellor, S. (2026). Opportunities and Barriers for AI Feedback on Meeting Inclusion in Socioorganizational Teams. Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26), pp. 1–17. https://doi.org/10.1145/3772318.3791135 (arXiv:2601.11750) - **Type:** conference — peer-reviewed: yes. The full text read is the arXiv manuscript ("Manuscript submitted to ACM"), which may differ slightly from the CHI version. - **Full text read:** yes (https://arxiv.org/pdf/2601.11750) - **Method:** Preregistered within-subjects lab study (control, then an Induced Hypocrisy Procedure (IHP) pre-meeting conversation with the AI agent), plus a 3-week within-subjects field study with semi-structured interviews. Outcomes: Davison meeting-success constructs, group attraction, speaking-time Gini, and mixed-effects models (p. 3, p. 8–10, p. 13 of PDF, §4, §6). - **n:** Lab: 28 participants in 7 groups of 3–5, 25-minute group tasks per condition. Field: 10 participants at a small consulting firm in recurring virtual meetings. Interviews with 9 of the 10 (mean 24 min) (p. 3, p. 9, p. 13 of PDF). - **Setting:** Virtual meetings with an AI agent "Emily" (GPT-4o-mini via the OpenAI API). Before the meeting she guides users through IHP goal-setting using speaking-time data. After the meeting she privately collects feedback and passes it on only with approval (p. 6–7 of PDF, §3). - **Findings:** 1. Lab: perceived meeting quality improved on 4 of 5 constructs, e.g. Discussion Quality median 0.88 to 1.00 (V = 4.5, p < 0.001, r = 0.69), Efficiency (p < 0.001, r = 0.59), Status Effects (p = 0.01, r = 0.50), Teamwork (p = 0.008, r = 0.44). Communication (r = 0.01) and group attraction (p = 0.11) did not change. Self-reported AI influence on participation: V = 5.5, p < 0.001, r = 0.85 (p. 11–12 of PDF, Table 1, §5). 2. Speaking-time balance: group-level Gini showed a moderate effect (r = 0.51) but was not significant (V = 22, p = 0.11, n = 7 groups). The mixed model gave β = −0.146 (χ2(1) = 2.08, p = 0.15). The authors describe a "likely small-to-moderate effect" (p. 11–12 of PDF, §5.1.2). The abstract's claim that speaking times became "more balanced" is therefore not statistically supported. 3. Field: no significant differences (e.g. AI influence p = 0.06, all meeting-quality constructs p ≥ 0.22). Barriers were time pressure, feedback prompts seen as misaligned, hierarchy (less upward feedback to seniors), and use for personal reflection rather than feedback exchange (p. 14–15 of PDF, Table 2, §7). - **Relevance:** (a) — | (b) LLM agent as a pre- and post-meeting feedback facilitator for meeting inclusion, tested in lab and field | (c) higher perceived discussion quality and efficiency in the lab, with no field effect - **Caveats:** Small n (7 lab groups, 10 field participants). Fixed order (control always first), with possible order and placebo effects, which the authors discuss. The speaking-balance effect is not significant. - **Screening check:** Meets criteria (empirical, groups of 3–5 humans in meetings with LLM agent involvement, method and n reported, English). #### C82 — LLM chat assistant improves divisive online political conversations - **Citation:** Argyle, L. P., Bail, C. A., Busby, E. C., Gubler, J. R., Howe, T., Rytting, C., Sorensen, T., & Wingate, D. (2023). Leveraging AI for democratic discourse: Chat interventions can improve online political conversations at scale. Proceedings of the National Academy of Sciences, 120(41), e2311627120. https://doi.org/10.1073/pnas.2311627120 - **Type:** journal — peer-reviewed: yes - **Full text read:** yes (PMC full text XML via Europe PMC, PMC10576030; page numbers not available in XML, so sections are cited) - **Method:** Preregistered randomized field experiment, randomized at the conversation level (one partner in treated chats gets GPT-3 rephrasing suggestions: restate, validate, politeness). Text analysis of tone and topic, plus a follow-up about 3 months later (§Results, §Materials and Methods). - **n:** 1,574 participants (October 2022) in dyadic chats between gun-policy opponents. 2,742 AI rephrasings were suggested, 1,798 (two-thirds) accepted. Only 698 (44%) were in chats that reached the full intended length (§Results "Treatment Effects"). - **Setting:** US national sample recruited through Bovitz, discussing gun regulation in anonymous text chat on a custom platform. GPT-3 acts as a real-time "moderator" that proposes rephrasings before a message is sent, and the user can accept, edit or ignore them (§Materials and Methods). - **Findings:** 1. Partners of participants who received ≥1 rephrasing reported significantly higher conversation quality (+4 percentage points) and democratic reciprocity (+2.5 pp). The reciprocity effect reached roughly 6 pp under full treatment (4+ rephrasings) (§Treatment Effects, Fig. 4). 2. Accepted rephrasings were 30% restate, 30% validate and 40% politeness. They scored higher on positive emotion, hedges, agreement and acknowledgment, but topic distribution did not change (χ2 = 0.150, N = 871, p = 1.00) (§AI Rephrasings: Tone and Topic). 3. The assistant produced no more policy-attitude change than control. The effects did not persist at about 3 months (§Treatment Effects, §Materials and Methods). - **Relevance:** (a) — | (b) LLM as a scalable real-time conversation moderator/coach in deliberative dialogue | (c) improves perceived conversation quality and reciprocity without shifting positions (decision content unchanged) - **Caveats:** Dyads rather than larger groups. Anonymous text chat. Effect sizes are small. Only 44% received full treatment (dose subgroups use placebo-controlled estimates). The effects are not durable. - **Screening check:** Meets criteria (empirical RCT, 2 humans interacting with LLM mediation, method and n reported, English). #### C773 — ModiBot: LLM moderator in normal and emotionally challenging focus groups - **Citation:** Zrubka, Z., Marczisák, B., Süli, P. P., Eigner, G., Gáspár, J., Benedek, M., & Sziklai, B. R. (2025). Piloting ModiBot: A Large Language Model-Based Moderator in Normal and Emotionally Challenging Focus Group Interactions. 2025 IEEE 25th International Symposium on Computational Intelligence and Informatics (CINTI), pp. 767–772. https://doi.org/10.1109/cinti67731.2025.11311748 - **Type:** conference — peer-reviewed: yes (IEEE symposium) - **Full text read:** abstract only (IEEE Xplore, no OA copy found) - **Method:** Pilot with four simulated focus groups (role-playing researchers) under scripted dynamics; quantitative evaluations plus qualitative sentiment and thematic analyses (abstract). - **n:** 4 focus groups (abstract); number of role-playing participants not stated — not verified. - **Setting:** Discord-based platform; ModiBot, customizable LLM-based moderator guiding onboarding, discussion, AI-mediated synthesis and reflective validation; scenarios: constructive discussion, anger, peer conflict, moderator-directed protest (abstract). Underlying LLM not verified. - **Findings:** 1. ModiBot was perceived as objective and able to ensure anonymity and structure (abstract). 2. It struggled to de-escalate emotional conflict, adapt to complex group behaviour and manage pacing; emotional flatness and rigid message flow reduced empathy and cohesion (abstract). 3. Data collection succeeded but social intelligence in group contexts was limited (abstract). - **Relevance:** (a) — | (b) LLM autonomously moderating a multi-person focus group, incl. conflict situations | (c) reduced empathy/cohesion under AI moderation - **Caveats:** Abstract only; role-playing researchers rather than real participants; 4 groups; no numbers available. - **Screening check:** Meets criteria on abstract — empirical pilot, 4 human focus groups moderated by an LLM, method and group count stated, English. Participant count not verified. #### GS4 — Focus Agent: LLM-moderated virtual focus groups - **Citation:** Zhang, T., Zhang, X., Cools, R., & Simeone, A. L. (2024). Focus Agent: LLM-Powered Virtual Focus Group. In Proceedings of the 24th ACM International Conference on Intelligent Virtual Agents (IVA '24), 1–10. https://doi.org/10.1145/3652988.3673918 - **Type:** conference — peer-reviewed: yes - **Full text read:** yes (https://arxiv.org/pdf/2409.01907) - **Method:** User study: five AI-moderated focus groups with human participants (VR or desktop) on "digital well-being", each followed by a researcher-led meta focus group; plus LLM-only focus-group simulations on the same topic; thematic and content analysis of transcripts (§5–6, pp. 5–6 of PDF). Pilot with 4 volunteers (§4, p. 4–5). - **n:** 23 participants (11 VR, 12 desktop; mean age 30) in 5 groups of 3–6 (p. 5 of PDF). - **Setting:** KU Leuven; AI moderator built on ChatGPT (GPT-4 rejected for ~20× cost after pilot showed similar opinions, p. 3 of PDF), with Whisper-based multi-speaker speech-to-text and Google TTS, in Mozilla Hubs; sessions scheduled 60 min, actual mean 51 min (SD 13) (p. 5 of PDF). - **Findings:** 1. Content analysis yielded 39 unique codes from human focus groups vs 47 from simulations; most human opinions were also covered by AI participants, but AI opinions were more repetitive/common across sessions (§6.1, p. 6 of PDF). 2. In meta focus groups participants credited the agent with steering discussion and inviting contributions, but the "prevailing sentiment" leaned to dissatisfaction: repeated questions, perceived lack of understanding, bias (e.g., "demonising" screen use), interruptions of participants mid-answer (§6.2, p. 6 of PDF). 3. Authors advise against deploying the agent as sole moderator; suggest AI-generated summaries/questions be used by human moderators instead (§7.2, p. 8 of PDF). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: direct test of an LLM running live multi-person focus groups | (c) AI and group ideation/decisions: — - **Caveats:** Small n (23, 5 groups); no human-moderator comparison; single topic; recording issue in group 3 (transcript reconstructed from notes, p. 6 of PDF); analysis led by first author. - **Screening check:** Meets criteria — empirical user study with stated method and n; LLM moderates groups of 3–6 interacting humans; English. #### S2-308 — Read the Room or Lead the Room: socio-cognitive dynamics in human-AI teaming - **Citation:** Choi, J., Samadi, M. A., JaQuay, S., Park, S., & Nixon, N. (2025). Read the Room or Lead the Room: Understanding Socio-Cognitive Dynamics in Human-AI Teaming. arXiv:2510.09944v1 [cs.HC], 11 Oct 2025. https://doi.org/10.48550/arXiv.2510.09944 (the arXiv PDF states no venue; LAK venue in the TSV is from Semantic Scholar and not verified) - **Type:** preprint (possibly also a conference paper; not verified) — peer-reviewed: no (preprint read) - **Full text read:** yes (arXiv PDF, https://arxiv.org/abs/2510.09944; pdf3/S2-308.pdf; printed pages 1–11) - **Method:** Classroom experiment with random assignment to control (3-human teams) vs treatment (2 humans + 1 AI teammate); 20-min text-chat collaboration on the TRAIL platform; discourse analysed with LIWC and Group Communication Analysis (GCA); ANOVA + Tukey post-hoc with Benjamini-Hochberg correction; supplementary qualitative discourse analysis (§4, pp. 4–5; §6, p. 8). - **n:** 179 undergraduate students; 37 control teams and 37 treatment teams; three experiments across two undergraduate classes at a US university (two in person, one via Zoom) (§4.1.2, p. 4). Note: 37×3 + 37×2 = 185 human seats, not 179; the paper does not explain the difference. - **Setting:** Undergraduates; NASA Moon Survival and Winter Survival ranking tasks (rank 15 items); AI teammate = OpenAI GPT-4 API with custom memory, prompted to co-construct knowledge, be empathetic, friendly and adaptive; random nicknames hid identity (§4.1, p. 4). - **Findings:** 1. The AI teammate acted as a cognitive facilitator/leader: vs humans in treatment teams it used more Clout (75.85 vs 46.92), Analytic (41.67 vs 31.81), Affiliation, Drives, BigWords, we-pronouns (5.16 vs 2.25), future focus, and fewer I-pronouns (1.02 vs 4.53); all AI vs H_TRT p < 0.001 (Table 2, p. 6). Qualitatively, AI teammates "frequently acted as dominant facilitators, guiding, planning, and driving group decision-making", setting priorities and repeatedly prompting for agreement, with humans giving "shorter, often passive, affirming responses" (§6, p. 8). 2. The AI failed to "read the room": higher GCA Participation (0.07 vs -0.03) and Internal Cohesion (0.24 vs 0.08) but lower Newness (0.45 vs 0.58) and Communication Density (0.04 vs 0.35), all p < 0.001, i.e., repetitive and verbose contributions ("Socially Detached") (Table 2, p. 6; pp. 6–8). Text reports AI utterances were nearly three times longer, F(2,213) = 74.36, p < 0.001 (p. 6). 3. Humans working with AI used more social-process language than humans in human-only teams (Social 12.58 vs 9.64, Tukey p < 0.001) (Table 2, p. 6; p. 7); qualitatively they treated the AI as a "secretary", offloading summarising/ordering and deferring to its decisions, engaging less in reasoning themselves (§6, p. 8). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: autonomous GPT-4 teammate spontaneously took the facilitator/organiser role in small-group decision tasks | (c) AI and group ideation/decisions: AI dominated decision-making, humans deferred; task performance/decision quality not reported - **Caveats:** Preprint (short paper); no task-performance or accuracy outcomes reported despite a scoring-possible task; Table 2 "Avg Word Count" column values (AI 7.96 vs H_TRT 30.57) contradict the text's claim that AI utterances were ~3× longer, suggesting a column mix-up; n arithmetic inconsistency (179 vs 185 seats); student sample; 20-min text chat; qualitative analysis method not described. - **Screening check:** Meets criteria: empirical, groups of 2 humans + GPT-4 interacting, method and n reported, English. #### R543 — AI team members and decision quality (hidden profile) - **Citation:** Zercher, D., Jussupow, E., Benke, I., & Heinzl, A. (2025). How Can Teams Benefit From AI Team Members? Exploring the Effect of Generative AI on Decision-Making Processes and Decision Quality in Team–AI Collaboration. Journal of Organizational Behavior. https://doi.org/10.1002/job.2898 - **Type:** journal — peer-reviewed: yes - **Full text read:** abstract only (Wiley PDF blocked by Cloudflare; volume/pages not verified) - **Method:** exploratory experiment, three knowledge configurations; process models derived (abstract) - **n:** 215 individuals in 81 teams (abstract) - **Setting:** hidden-profile decision task; (1) human teams with asymmetric knowledge, (2) teams collaborating with a generative AI team member holding centralized knowledge, (3) teams with AI and asymmetric knowledge (abstract); specific AI model not verified - **Findings:** 1. Teams with centralized AI knowledge made more accurate decisions than human teams, attributed to reduced decision-making asymmetries, trust in AI, beneficial AI information processing and a balanced AI collaboration focus (abstract). 2. Teams with asymmetric AI knowledge showed only moderate reductions in decision-making asymmetries and did not outperform human teams (abstract). 3. New AI-specific asymmetries emerged — mistrust, non-beneficial AI information processing, critical AI collaboration focus (abstract). - **Relevance:** (a) — | (b) — | (c) a generative AI team member can improve group decision quality when it holds pooled information, but not when knowledge is distributed - **Caveats:** abstract only; exact accuracy figures and statistics not verified; "exploratory" design. - **Screening check:** meets criteria — experiment with multi-human teams collaborating with generative AI on a group decision, method and n stated, English. #### R173 — AI-assisted vs AI-led vs human-only teams (reproducibility) - **Citation:** Brodeur, A., Valenta, D., Marcoci, A., Aparicio, J. P., Mikola, D., Barbarioli, B., et al. (2026). AI-assisted teams outperform AI-led teams but not human-only teams in assessing research reproducibility in quantitative social science. Proceedings of the National Academy of Sciences, 123(22), e2524747123. https://doi.org/10.1073/pnas.2524747123 - **Type:** journal — peer-reviewed: yes - **Full text read:** abstract only (PNAS PDF blocked by bot protection; not in Europe PMC as OA) - **Method:** randomised experiment with three team conditions: human-only, AI-assisted (ChatGPT as collaborative tool), AI-led (ChatGPT with minimal human oversight) (abstract) - **n:** 288 researchers randomly assigned to 103 teams (abstract) - **Setting:** researcher teams reproducing published results from leading social science journals, detecting coding errors and proposing robustness checks; ChatGPT (model version not verified) - **Findings:** 1. Reproduction rates: human-only 94% vs AI-assisted 91% (comparable); AI-led 37% (abstract). 2. Human-only teams identified significantly more major coding errors than AI-assisted teams; otherwise the two performed similarly on most outcomes (abstract). 3. AI-led teams detected fewer errors in all categories, proposed weaker robustness checks and required more time; AI assistance "provided no measurable advantages" (abstract). - **Relevance:** (a) — | (b) — | (c) LLM assistance did not improve expert team judgement quality and was associated with fewer major errors detected; autonomous AI-led work was much worse - **Caveats:** abstract only (effect sizes, team size distribution, model details not verified); task is technical verification work, not a facilitated session. - **Screening check:** meets criteria — randomised experiment, multi-person human teams working with an LLM, method and n stated, English. #### GS2 — Undetected AI personas in human–AI triads - **Citation:** Yan, L., Han, X., Zhang, Y., Greiff, S., Molenaar, I., Fan, Y., Martinez-Maldonado, R., Zhao, L., Li, X., Jin, Y., & Gašević, D. (2025). The Social Blindspot in Human–AI Collaboration: How Undetected AI Personas Reshape Team Dynamics. arXiv:2512.18234. https://arxiv.org/abs/2512.18234 - **Type:** preprint — peer-reviewed: no (preprint; "submitted to Elsevier") - **Full text read:** yes (https://arxiv.org/pdf/2512.18234) - **Method:** Large-scale mixed-design online experiment, 3 tasks (analytical survival ranking, ethical dilemma, creative writing) × 6 group conditions (3H; 2H+1AI supportive/contrarian; 1H+2AI supportive/contrarian/mixed); individual–group–individual sequence with a 10-minute synchronous text chat; participants not told AI might be present; mixed models with multiplicity correction; linguistic mediation analysis (§3.2, pp. 4–5; §3.10–3.12). - **n:** 1,047 entered; final N = 905 participants in 572 three-person groups (Prolific); N = 871 for detection analyses (§3.1, pp. 3–4). Per-condition counts in Table 2 (p. 5). - **Setting:** Adults (18–65, English-fluent) on a custom chat platform; LLM-based AI teammates with scripted supportive (affiliative, consensus-building) or contrarian (challenging, dissent-oriented) personas, forbidden to reveal AI identity; LLM model not named in main text (not verified) (§3.4–3.5, pp. 7–8). - **Findings:** 1. Low detectability: AI teammates correctly identified 30.8% (95% CI 27.7–33.9%); humans misclassified as AI in 27.6% (§4.2, p. 15). 2. Group discussion quality: 2 humans + 1 supportive AI scored higher than all-human teams, β = 0.590, 95% CI [0.259, 0.922], p_adj = .012; 2 humans + 1 contrarian AI scored lower, β = −0.856, 95% CI [−1.257, −0.454], p_adj = .001 (§4.3, p. 16). 3. Psychological safety fell with contrarian personas: H2_C β = −0.671, p_adj = .003; H1_C β = −0.881, p_adj = .049; supportive personas n.s. (H2_S β = 0.152, p_adj = 1.00). Individual performance gains generally n.s., except the analytical task where H2_C outperformed all-human teams (β = 0.88, p_adj = .012) (§4.3–4.4, p. 16). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: a supportive (facilitator-like, consensus-building) AI member raised discussion quality | (c) AI and group ideation/decisions: AI persona shapes discussion quality and psychological safety in small-group chat, including a creative task - **Caveats:** Preprint; deceptive design (AI undisclosed); short 10-min online text chats with crowdworkers; AI is a peer member, not an explicit facilitator; 1H+2AI conditions contain only one human. - **Screening check:** Meets criteria — randomized experiment with method and n reported; 2H+1AI and 3H conditions have ≥2 humans interacting as a group with an LLM agent; English. #### C801 — SeeSawBot: LLM mediator across private DMs and team Slack channels - **Citation:** Wang, Y., Lei, K., Chiu, S.-Y., Isbister, K., Lee, D., & Ringland, K. E. (2026). SeeSawBot: An LLM-Driven Chatbot Mediating Across Private and Shared Slack Channels to Support Team Dynamics. Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26), Barcelona, 23 pages (pp. 1–23). https://doi.org/10.1145/3772318.3791880 - **Type:** conference — peer-reviewed: yes - **Full text read:** yes (author copy, https://wongyihe.github.io/assets/pdf/Wang%20et%20al.%20-%202026%20-%20SeeSawBot%20An%20LLM-Driven%20Chatbot%20Mediating%20Across%20Private%20and%20Shared%20Slack%20Channels%20to%20Support%20Team.pdf) - **Method:** Research-through-design technology probe: formative mock-up study (group interview + role-play), then 8-week field deployment with iteratively revised bot versions; bi-weekly open-ended reflection surveys, chat logs, post-deployment semi-structured interviews; thematic analysis (§3–6). - **n:** Formative study n = 10 (p. 5, §4). Deployment: 139 students enrolled, 105 consented, 18 teams of 5–6 members (p. 8, §6.1); survey waves n = 63, 51, 57, 52; interviews n = 9 (p. 2, §1). - **Setting:** 10-week undergraduate information systems course (US university) with two team projects; SeeSawBot built on GPT-4o, in team Slack channels and DMs: weekly reminders, weekly DM recaps interpreting group dynamics, nudges to quiet members, private check-ins that could be escalated to public prompts (pp. 6–9, §5–6). - **Findings:** 1. Private-to-public synthesis: students used DMs to "workshop" thoughts privately before bringing them to the team; in conflicts the bot DM'd each member for their perspective and then prompted open discussion in the channel (p. 10, §7.1.1). 2. Participation balancing and social buffering: the bot pinged non-contributing members and students delegated socially risky reminders to it ("so people don't have to be annoying"), redistributing emotional labour in leaderless teams (p. 12, §7.2.2–7.2.3). 3. Roles shifted across team development stages; during storming the bot surfaced tensions, though its private check-ins "often receive no response"; reminders and summaries became informal stand-up routines in norming (pp. 13–14, §7.3, Fig. 7). - **Relevance:** (a) — | (b) LLM agent acting as an ongoing team mediator/facilitator (nudging participation, synthesising private input for the group, recaps) | (c) qualitative effects on participation balance and conflict surfacing; no decision-quality measures - **Caveats:** Qualitative only, no comparison condition; student teams in one course; bot changed during deployment; only 9 interviews; asynchronous Slack teamwork rather than facilitated meetings. - **Screening check:** Meets criteria — empirical field deployment with 18 multi-person teams interacting with a GPT-4o mediator; group processes measured; method and n reported; English. Classroom setting, but the AI mediates team group processes rather than individual study, so not EX-EDU. #### S2-479 — "Shall We Dig Deeper?": LLM agent intervention styles in asynchronous group discussions - **Citation:** Zhang, Y., Li, W., Wang, X., Yuan, K., Ma, S., & Ma, X. (2026). "Shall We Dig Deeper?": Designing and Evaluating Strategies for LLM Agents to Advance Knowledge Co-Construction in Asynchronous Online Discussions. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26), Barcelona, Spain. ACM, 27 pages. https://doi.org/10.1145/3772318.3790551 (also arXiv:2509.23327v2) - **Type:** conference — peer-reviewed: yes (CHI '26). The copy read is arXiv v2 (1 Feb 2026); page 14 of that PDF still carries "Manuscript submitted to ACM" residue. - **Full text read:** yes (local PDF pdf3/S2-479.pdf = arXiv 2509.23327v2, https://arxiv.org/abs/2509.23327). No printed page numbers, so pages below are "p. X of PDF". - **Method:** (1) A 3-hour online design workshop (12 participants: 3 AI designers and 9 frequent online-discussion contributors, working in subgroups of 4) derived phase sufficiency criteria and task- vs relationship-oriented intervention strategies (Telling / Selling / Participating / Delegating) (p. 5 of PDF, §3.2). (2) A mixed-methods within-subject lab study with five conditions (four agent styles plus a human-only baseline), one discussion thread per day over five consecutive days, counterbalanced by balanced Latin squares (60 unique sequences). Measures: thread-level coding of knowledge co-construction phases by two researchers, 7-point in-task surveys, NASA-TLX subscales, IOS, and semi-structured interviews. Analysis used Friedman and Wilcoxon tests with Bonferroni correction, GLMMs with BH correction, and thematic analysis (pp. 9–13 of PDF, §4). - **n:** Main study N = 60 (24 female, 36 male, aged 19–33; active contributors to asynchronous discussion platforms). An a priori power analysis gave a minimum of 31 (p. 11 of PDF, §4.2). **Group size: 6 participants per thread**: "Pilot study showed that groups of six participants per thread yielded sufficient interaction… Each thread is discussed by a distinct group of 6 participants" (p. 9 of PDF, §4.1). 50 discussion threads in total (p. 9 of PDF). 10 threads per condition is implied by 50/5 and by the percentages in steps of 10% in Table 3, but not stated explicitly. Group membership was reshuffled each day ("no two participants followed the exact same trajectory, preventing fixed subgroups", p. 12 of PDF). The limitations section, however, says each thread ran "within a fixed six-person group over five consecutive days" (p. 23 of PDF, §7), which is inconsistent. Design workshop: 12 participants (p. 5 of PDF). - **Setting:** A web-based Discourse-like forum platform built for the study. Participants posted asynchronously in a one-day window per thread on five open-ended, non-specialist topics (e.g., "What's the most effective way to learn a new language"). The LLM agent detected the current discussion phase and sufficiency and intervened in one of four styles. Model: "During our experiment (August 2025), we used Gemini 2.5 Flash" (footnote 2, p. 8 of PDF, §3.3.2). Setting country is not verified. Participants were compensated at the "local standard hourly rate" (p. 12 of PDF). - **Findings:** 1. Agent presence deepened discussion. Max Phase Reached (mean ± SD): Baseline 2.10 (0.32), Telling 2.80 (0.63), Selling 2.90 (0.88), Participating 2.60 (0.52), Delegating 2.20 (0.42). Phase 2 sufficiency: 10% baseline vs 80% / 60% / 70% / 20%. Phase 3 sufficiency: 0 baseline vs 30% for Telling, Selling and Participating. Phase 4 sufficiency: 0 in every condition (Table 3, p. 14 of PDF). The pairwise GLMM contrasts vs baseline were only marginal (e.g., Max Phase Baseline–Telling p = 0.083, OR 0.05; Phase 2 Baseline–Telling p = 0.052, OR 0.03) (Table 4, p. 15 of PDF). The authors call these "practically meaningful improvements despite not meeting conventional thresholds" (p. 14 of PDF, §5.2.1). 2. Agents crowded out human–human exchange. Human–human replies per thread: Baseline 6.40 (2.32), Telling 2.90 (3.81), Selling 3.90 (2.69), Participating 4.10 (2.60), Delegating 6.10 (3.41) (Table 3, p. 14 of PDF). Some participants disengaged when the agent's viewpoints were well formed: "It felt so smart that I didn't need to add anything" (P21) (p. 15 of PDF). 3. Style mattered for acceptance. The Participating style received the most agent-comment likes (2.20 vs Telling 0.90; Participating–Telling p = 0.048, RR 2.44). Delegating was rated significantly less helpful and appropriate, and "Most participants (44/60) reported largely ignoring its interventions". Selling raised Mental Demand significantly relative to baseline. Direct requests to summarise or reflect "were largely ineffective in transitioning discussions into the co-construction stage" (pp. 15–18 of PDF, §5.2–5.3). Manipulation check: 82.5% of style classifications matched the intended style, χ²(9) = 432.58, p < .001 (p. 13 of PDF, §5.1). - **Relevance:** (a) — | (b) An LLM agent acting as an automated discussion facilitator in asynchronous 6-person groups, with a controlled comparison of facilitation styles vs no facilitator | (c) Deeper knowledge co-construction with task-oriented or participating agents, but fewer human–human replies and some disengagement. - **Caveats:** Key contrasts vs baseline are only marginally significant, with 10 threads per condition implied. Young sample (19–33). Simulated one-day asynchronous threads on lay topics rather than a real workshop or meeting. The text is inconsistent about whether groups were fixed or reshuffled (p. 12 vs p. 23 of PDF). The citation year in s2_decisions.tsv (2025) differs from the paper (CHI '26, 2026). - **Screening check:** Meets the criteria. It is empirical, with groups of 6 humans interacting in discussion with an LLM agent performing a facilitation function, and it reports method and n (N = 60). English. Note that the discussions are asynchronous online forums, not synchronous meetings. #### R1 — PeerGPT (LLM peer agent as team moderator vs participant) - **Citation:** Liu, J., Yao, Y., An, P., & Wang, Q. (2024). PeerGPT: Probing the Roles of LLM-based Peer Agents as Team Moderators and Participants in Children's Collaborative Learning. Extended Abstracts of the CHI Conference on Human Factors in Computing Systems (CHI EA '24), ACM, pp. 1-6. https://doi.org/10.1145/3613905.3651008 - **Type:** conference (late-breaking work) — peer-reviewed: yes (extended abstract) - **Full text read:** yes (arXiv, https://arxiv.org/pdf/2403.14227; pages = arXiv PDF pages) - **Method:** exploratory pilot study; two design-based collaborative learning workshops (between-group: agent as moderator vs agent as participant); qualitative thematic analysis of transcripts (MAXQDA) plus word-segment counts (§3, p. 2-3) - **n:** 6 children aged 11-12 in 2 groups of 3 (p. 2, Table 1 p. 3); 262 minutes of audio/video; 583 segments (Group A) and 612 segments (Group B) analysed (§4, p. 3) - **Setting:** children's workshop on "urban and environmental design challenges in 2050"; peer agent "Blue" built on OpenAI GPT-3.5 with voice output on a tablet; ~2 hours per workshop (p. 2-3) - **Findings:** 1. As moderator, the agent could "relatively effectively control the pace and structure of the discussion" but lacked authority, so children sometimes ignored or questioned its instructions; its longer messages often failed to capture attention (p. 2; §4.1.1, p. 3). 2. As moderator, children had fewer opportunities for free discussion among themselves and the agent could not intervene promptly when children expressed negative emotions toward peers (§4.2.1, p. 5). 3. As participant, the agent's contributions stimulated creative thinking and were used by the child moderator to steer discussion; its participation was concentrated in the first two workshop stages and diminished in prototyping/presentation, where communication became non-verbal (§4.1.2, §4.3, p. 4-5). - **Relevance:** (a) — | (b) LLM acting as group moderator/facilitator in a structured workshop: manages structure but lacks authority | (c) LLM participant can stimulate group ideation - **Caveats:** very small n (6 children, 2 groups, one per condition), children, education context, late-breaking work; no quantitative outcome comparison. - **Screening check:** meets criteria — empirical (qualitative), groups of 3 children interacting with an LLM agent that facilitates a structured group session, group-level conversation analysed, method and n stated, English. #### C327 — LLM-driven embodied virtual agent as mediator in triadic XR decision-making - **Citation:** Han, B., Dong, Z., Zhang, J., Wen, R., Hirai, T., Clark, A., & Piumsomboon, T. (2026). Exploring Mediation by an Embodied Virtual Agent in Immersive Triadic Collaborative Decision-Making. IEEE Transactions on Visualization and Computer Graphics, 32(5), 4451–4460. https://doi.org/10.1109/TVCG.2026.3679103 - **Type:** journal (IEEE TVCG) — peer-reviewed: yes - **Full text read:** abstract only (IEEE paywall; no OA copy found in PMC/Europe PMC; a related University of Canterbury thesis record "Triadic human-agent dynamics in collaborative decision-making within extended reality" exists but was blocked by Cloudflare and not read) - **Method:** User study of an XR system comparing agent presence and embodiment (conditions implied by abstract: embodied agent, disembodied agent, and presumably no agent) with quantitative measures (co-presence, perceived collaboration, preference, efficiency, satisfaction) and qualitative analysis; exact design (within/between) not verified. - **n:** "pairs (n=24)" per abstract — unclear whether 24 pairs or 24 participants; not verified. - **Setting:** Two users with conflicting goals collaborating on an office design task in XR with an LLM-driven Embodied Virtual Agent as mediator; LLM model not verified. - **Findings:** 1. Agent embodiment significantly enhanced co-presence, which fostered interactions leading to better perceived collaboration and higher user preference; frequent interactions with the disembodied agent were associated with lower preference (abstract; statistics not verified). 2. Agent presence did not improve task efficiency or satisfaction (abstract). 3. Qualitative analysis showed the LLM agent spontaneously adopted facilitative and evaluative mediation strategies aligned with collaborating and compromising conflict-resolution modes, without explicit programming (abstract). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: LLM agent acting as mediator/facilitator in a live two-person negotiation | (c) AI and group ideation/decisions: no efficiency/satisfaction gain; value mainly in social dynamics and perceived collaboration - **Caveats:** Abstract only; group = dyad plus agent; sample size ambiguous; XR lab task. - **Screening check:** Meets criteria on abstract (empirical user study, ≥2 humans with an LLM mediator, n reported, English) — details to verify from full text. #### OA2160 — Cognitive and social awareness in LLM-chatbot-assisted group collaboration - **Citation:** de Jong, S., Wester, J., Schrills, T., Secher, K. S., Griggio, C. F., & van Berkel, N. (2024). Assessing Cognitive and Social Awareness among Group Members in AI-assisted Collaboration. In Proceedings of the International Conference on Mobile and Ubiquitous Multimedia (MUM '24), pp. 338–350. https://doi.org/10.1145/3701571.3701582 - **Type:** conference — peer-reviewed: yes (ACM; venue name inferred from the DOI prefix 3701571 and not verified) - **Full text read:** abstract only (ACM DL and the AAU repository PDF were blocked by bot protection, HTTP 403; abstract taken from OpenAlex) - **Method:** "In-situ study": a distributed text-based group task, followed by a survey and semi-structured interviews (abstract). - **n:** N = 48 participants in 16 groups of 3 (abstract). - **Setting:** Travel-planning task in which each group member had conflicting goals; each group chat included an LLM-based chatbot that could be asked for help. The specific LLM is not verified. - **Findings:** 1. Having the chatbot present "helped increase group awareness as users are forced to clearly and transparently formulate their intentions when prompting the chatbot" (abstract). 2. The chatbot's compromise suggestions based on the chat history "helped participants reach a consensus" (abstract). 3. Effect sizes and survey statistics: not verified (full text not accessed). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: an LLM chatbot as a mediator inside small-group chat | (c) AI and group ideation/decisions: compromise suggestions supported reaching consensus - **Caveats:** Abstract only; small n; lab-style task; the "in-situ" meaning is unclear. - **Screening check:** PASSES (empirical; groups of 3 humans interacting with an LLM chatbot; method and n reported in the abstract). #### OA1518 — GenAI as teammate: social loafing in human–AI groups (PACIS 2026) - **Citation:** Matussek, M., Wendland, R., & Hendriks, P. (2026). Generative AI as a teammate: Exploring social loafing and individual effort in human–AI teams. PACIS 2026 Proceedings (General topic, paper 9). https://aisel.aisnet.org/pacis2026/general_topic/general_topic/9 (the batch metadata names "Journal of the Association for Information Systems" as venue; the AIS eLibrary page lists it as PACIS 2026 Proceedings, online 2026-06-12) - **Type:** conference — peer-reviewed: unclear (PACIS is a refereed AIS conference, but review status of this paper not verified) - **Full text read:** abstract only (PDF at aisel.aisnet.org/cgi/viewcontent.cgi?article=1248&context=pacis2026 blocked by a bot-protection page; abstract taken from the AIS eLibrary landing page) - **Method:** per abstract: group experiment with a brainstorming and evaluation task on a text-based platform (Discord) with an LLM chatbot, analysed with quantitative and qualitative data ("qualitative findings" mentioned); exact design, measures and statistics not verified. - **n:** 28 participants in 7 groups of 4 (abstract). - **Setting:** "validated brainstorming and evaluation task" on Discord with assistance from an LLM chatbot (model not stated in abstract); authors affiliated with TU Darmstadt. - **Findings:** 1. "AI does not inherently increase social loafing" (abstract). 2. "While some participants offloaded effort onto the AI, most used it strategically to extend and refine their own ideas" (abstract). 3. Qualitative findings indicate AI changes mediators of social loafing such as responsibility diffusion and group cohesion, depending on how it is perceived and engaged with (abstract). No numbers available. - **Relevance:** (a) — | (b) LLM chatbot embedded in small-group brainstorming/evaluation on a chat platform | (c) effort and social loafing in AI-supported group ideation - **Caveats:** abstract only; small n (28 / 7 groups); no statistics verifiable; unclear whether there was a no-AI comparison condition; venue mismatch in batch metadata (JAIS vs. PACIS). - **Screening check:** likely meets criteria (empirical, n reported, groups of 4 humans with an LLM), but method details not verified from full text. #### C53 — tAIfa: AI-generated automated feedback for team effectiveness and cohesion - **Citation:** Almutairi, M., Chiang, C., Bai, Y., & Gómez-Zará, D. (2025). tAIfa: Enhancing Team Effectiveness and Cohesion with AI-Generated Automated Feedback. In Proceedings of the 4th Annual Symposium on Human-Computer Interaction for Work (CHIWORK '25), 25 pages. https://doi.org/10.1145/3729176.3729197 - **Type:** conference (full paper) — peer-reviewed: yes - **Full text read:** yes (arXiv version, https://arxiv.org/pdf/2504.14222v1) - **Method:** Preregistered between-subjects laboratory experiment (teams randomized to no feedback vs. LLM-generated individual + team feedback after each of three tasks); transcript metrics, task performance vs. expert rankings, post-task surveys, NASA-TLX, repeated-measures ANOVA; preceded by a prompt-evaluation survey with Prolific raters (§3.3, §5, p. 5–10 of PDF). - **n:** Main study 54 participants in 18 three-person teams (10 treatment, 8 control); 39 university participants (mostly undergraduates) + 15 Prolific (§5.1, p. 8–9). Prompt-evaluation study: 30 recruited, 26 analysed (§3.3, p. 5–6). - **Setting:** Online teams collaborating via Slack text chat on three "survival" ranking tasks (10 min each); tAIfa Slack bot combines rule-based communication metrics with LLM analysis and delivers private individual and public team feedback; model GPT-o1-preview via Microsoft Azure OpenAI (§4.5, p. 8). - **Findings:** 1. Treatment teams talked longer (Task 3: M = 8.80 vs 6.50 min, t(49.35) = −2.11, p < 0.05; RM-ANOVA main effect F(1,154) = 4.21, p < 0.05, d = 0.57) and had more speaker turns per member (Task 3: 7.10 vs 5.36, t(50.52) = −2.03, p < 0.05, d = 0.55; F(1,154) = 5.74, p < 0.05) (§6.1, p. 10). 2. Task performance did not differ significantly (Task 3: 56.7% vs 45.8%, t(15.80) = −0.93, p > 0.05); self-reported efficiency, interaction quality, cohesion, satisfaction and viability showed no significant differences (cohesion slightly lower with tAIfa: M = 4.16 vs 4.23) (§6.1, p. 10). 3. Perceived effort (NASA-TLX) was higher with tAIfa (M = 4.00 vs 3.32; condition significant in two-way RM-ANOVA, F(1,149) = 15.36, p < 0.05) (§6.1, p. 11). - **Relevance:** (a) — | (b) LLM agent giving process feedback to teams between task rounds (a coaching/facilitation function) | (c) more participation and balanced turn-taking, but no significant gain in decision quality or cohesion - **Caveats:** Small n (18 teams), mostly student sample, text-only lab tasks; despite title, cohesion was not significantly improved. - **Screening check:** Meets criteria — randomized experiment with 3-person teams receiving LLM-generated feedback during group decision tasks; method and n reported. #### R105 — Proactive Conversational Agents with Inner Thoughts - **Citation:** Liu, X. B., Fang, S., Shi, W., Wu, C.-S., Igarashi, T., & Chen, X. A. (2025). Proactive Conversational Agents with Inner Thoughts. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI '25), ACM, pp. 1-19. https://doi.org/10.1145/3706598.3713760 - **Type:** conference — peer-reviewed: yes - **Full text read:** yes (arXiv, https://arxiv.org/pdf/2501.00383; pages = arXiv PDF pages) - **Method:** multi-part: GPT next-speaker-prediction benchmark on a multi-party chat corpus (§3); formative study of human group chats with think-aloud review (§4); technical evaluation of 100 simulated AI-only conversations rated by humans (§6); user study with human pairs chatting with the proactive AI on Slack, within-subjects over three proactivity styles (§7) - **n:** formative study 24 participants in groups of three (p. 5); human evaluation 10 evaluators rating 100 simulated conversations (p. 12); user study 12 participants in 6 pairs (§7.2, p. 14) - **Setting:** text-based casual group chat; LLM agents (GPT-3.5/GPT-4 family; fine-tuned GPT-3.5 for baseline) implementing an "Inner Thoughts" framework that decides when to speak by intrinsic motivation (p. 4, p. 11) - **Findings:** 1. In human ratings of simulated conversations, Inner Thoughts beat next-speaker prediction on all 7 metrics, strongest for turn appropriateness (U = 577.0, p = 2.4×10^-6) and coherence (U = 636.0, p = 1.6×10^-5); evaluators preferred Inner Thoughts conversations 82% of the time (§6.6, p. 12). 2. In the user study (human pairs + AI), the "Active contributor" setting was most preferred (6 participants), "Non-stop chatter" 4, "Selective participant" 2 (7 rated it worst); participants noted the active AI initiated new topics when conversation dried up and invited others by name (§7.3.2-7.3.3, p. 15). 3. Ratings across the three styles were mostly not significantly different; Selective Participant was rated lower than Non-stop Chatter on social presence and extroversion (p < 0.05) (§7.3.2, p. 15). - **Relevance:** (a) — | (b) design of when an LLM should speak in multi-party conversations (turn-taking/proactivity relevant to AI co-facilitators) | (c) — - **Caveats:** casual chat, not a facilitated or task-oriented session; the main quantitative evaluation is on AI-only simulated conversations (EX-SIM component); user study small (6 pairs), institutional sample. - **Screening check:** meets criteria via the user study (two humans interacting as a group with an LLM agent; method and n stated). The simulation component alone would fail (EX-SIM). #### S2-286 — Ethics Training Agents: LLM facilitator + LLM persona participants in human group ethics discussions - **Citation:** Seo, Y., Jang, S., Park, H., Gutierrez, R. S., Seering, J., & Lee, U. (2026). Ethics Training Agents: Facilitating Group-Based Ethics Education with Role-Playing and Discussion for Ethical Reflection and Exploration. arXiv preprint arXiv:2609.11529v1 [cs.HC] (10 Sep 2026). https://arxiv.org/abs/2609.11529 - **Type:** preprint — peer-reviewed: no (preprint) - **Full text read:** yes (arXiv 2609.11529v1 PDF, pdf3/S2-286.pdf; printed page numbers) - **Method:** Single-arm user study with pre–post measurement and no control condition. Ethical sensitivity was compared pre vs post with paired t-tests and Wilcoxon tests. Peer ratings of humans vs LLM agents used Kruskal–Wallis. Discussion logs gave the question-addressing ratio, tested at group level with Wilcoxon. Semi-structured focus-group interviews were analysed by reflexive thematic analysis (§5.4, pp. 13–14). - **n:** 45 undergraduate STEM students (freshmen excluded; 42% female, 58% male; age M = 22.1, SD = 1.7) in 15 groups of 3 humans. Each group discussed with 3 LLM persona participants (care ethics, deontology, pragmatic) plus an LLM facilitator (a "3:3" setting). One session per group of about 2 h, including about 60 min of multi-stage discussion (p. 12, §5.1–5.2; p. 16). - **Setting:** An offline university classroom (authors at KAIST, Korea; payment in KRW; IRB-approved) using a web chat system. It is a multi-agent digital adaptation of the Judgment Call ethics workshop game, with a Black Mirror smart-lens scenario in which participants act as engineers. GPT-4o implements all agents via LangGraph. The LLM facilitator handles flow control (stacking/hand-raising with random speaking order, designating answerers, time management), summarization and idea selection. The discussion structure follows Kaner et al.'s divergence–groan–convergence (§3, pp. 5–8; §4; §5.3). A trained human facilitator was on standby to intervene (p. 11). - **Findings:** 1. Ethical sensitivity rose from pre to post: issue awareness Cohen's d = 1.43 and consequence reasoning d = 1.28, both p < .001; Wilcoxon robustness check r = .99 and .98 (p. 14, §6.1). There was no control group, and the authors note novelty and general learning effects cannot be ruled out (p. 23, Limitations). 2. Participants said the LLM facilitator kept pace and coherence through turn control, time management and summaries, which reduced cognitive burden. Some found its strict turn restrictions uncomfortable and asked for partial interruptions (pp. 15–16, §6.2.2). The stacking format allowed many ideas in parallel but limited in-depth discussion of any single idea (p. 15, §6.2.1). 3. Human peers were rated significantly higher than every LLM persona on contribution, diversity and influence (Kruskal–Wallis, p < .05 for all pairwise comparisons). Participants attributed this to sycophantic, overly accepting agents that did not show their reasoning (pp. 17–18, §6.3). Even so, 79.7% of questions (59 of 74) went to LLM agents rather than humans, against 60% expected from availability (Wilcoxon vs .60, M = .81, SD = .29, p = .017, r = .72). Human-to-human exchange fell; participants reported lower social pressure (pp. 19–20, §6.4.1). - **Relevance:** (a) facilitator use of AI: — (the AI replaces the human facilitator; no facilitators studied) | (b) AI planning/running/supporting group sessions: an LLM facilitator runs a structured, workshop-game-based group session (turn-taking, timing, summaries, idea selection), with participant accounts of what worked | (c) AI and group ideation/decisions: LLM persona participants broadened the range of perspectives but reduced depth and human–human interaction; participants rated their contributions below humans' - **Caveats:** Preprint. Single-arm pre–post design with no control condition. Student sample in one Korean university, one scenario, one session per group. The ethical-sensitivity instrument was designed by the authors. A human facilitator was on standby. - **Screening check:** Meets criteria. The paper is empirical, with groups of 3 humans interacting with GPT-4o agents, including an LLM that facilitates the structured group session. It reports its method and n (45 participants / 15 groups) and is in English. #### S2-290 — Piecing Together Teamwork: LLM-based educational jigsaw agent - **Citation:** Doherty, E. J., Perkoff, E. M., von Bayern, S., Zhang, R., Dey, I., Bodzianowski, M., Puntambekar, S., & Hirshfield, L. (2025). Piecing Together Teamwork: A Responsible Approach to an LLM-based Educational Jigsaw Agent. In Proceedings of the CHI Conference on Human Factors in Computing Systems (CHI '25), pp. 1–17. https://doi.org/10.1145/3706598.3713349 (author list and pages 1–17 from OpenAlex/Semantic Scholar metadata; not verified against the paper itself) - **Type:** conference — peer-reviewed: yes (CHI full paper; per metadata) - **Full text read:** abstract only (ACM DL served a Cloudflare bot challenge; one quick arXiv title search found no preprint) - **Method:** Human-centered design plus evaluation; groups assigned to be assisted by the LLM agent or not (between-groups comparison per abstract). Assignment procedure (randomization or not) not verified. - **n:** 145 participants aged 12–17 in 58 groups (dyads and triads) (abstract). Split between agent vs. no-agent conditions not verified. - **Setting:** Middle- and high-school classrooms; small-group jigsaw collaborative activity; LLM-powered conversational agent that intervenes in collaborative conversations "when a teacher is unable to assist the students" (abstract). Specific LLM not verified. - **Findings:** 1. Groups with the agent showed decreased self-reported ratings of social loafing compared to groups without the agent (abstract; effect sizes/p-values not verified). 2. Groups with the agent showed increased use of language related to respectful collaboration in interactions with the agent compared to those without (abstract; not verified in full text). 3. The paper frames the design as a "responsible approach", addressing ethical concerns of LLM use with young populations (abstract). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: LLM agent acting as small-group collaboration facilitator in place of an absent teacher | (c) AI and group ideation/decisions: group process outcomes (social loafing, respectful collaboration language), not decision quality - **Caveats:** Abstract only; no numbers beyond n verifiable; school-age sample (12–17); outcomes are self-report and language measures; exact statistics, LLM model and condition sizes unknown. - **Screening check:** Meets criteria on the abstract: empirical, groups of 2–3 humans collaborating with an LLM agent facilitating, method and n reported, English. Full-text confirmation not possible. #### S2-251 — Social robot Haru assisting dynamic group discussion (LLM dialogue + autonomous gaze) - **Citation:** Tang, F., Hu, M., Fang, Y., Yu, H., Nichols, E., Gómez, R., & Li, G. (2025). Social Robot Haru Assisting Dynamic Group Discussion with Autonomous Eye Gaze Behavior. In 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 6372–6379. https://doi.org/10.1109/IROS60139.2025.11245910 (author list per OpenAlex metadata; the seventh author is Guangliang Li; venue name from the DOI prefix and TSV, not verified against the paper) - **Type:** conference — peer-reviewed: yes (IROS; not verified from full text) - **Full text read:** abstract only (IEEE Xplore closed access; OpenAlex reports no open-access copy; a quick arXiv title search found nothing) - **Method:** System development (three modules: an LLM-based Dialogue Assistance module; a Balancing and Welcoming Behavior module with verbal behaviors to raise engagement and welcome newcomers; an Autonomous Eye Gazing module trained with deep reinforcement learning in simulation and transferred to the physical robot), followed by a user study. Study design, conditions and measures are not verified. - **n:** 50 subjects (per abstract). Group sizes, number of groups/sessions and conditions not verified. - **Setting:** Social robot Haru embedded as an "embodied chatbot" through integration with large language models (which LLM is not stated in the abstract) in dynamic multi-party group discussions where users can join mid-discussion. Location and task not verified. - **Findings:** 1. The authors report that the user study with 50 subjects shows "the significant performance of our system in assisting dynamic group discussion" (abstract). No effect sizes, statistics or comparison conditions are available in the abstract. 2. The system is designed to balance participation by gazing at, and verbally encouraging, less-engaged users, and to welcome newcomers (abstract; design description, not an outcome). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: an LLM-powered robot moderates multi-party discussion (balancing participation, welcoming newcomers); outcomes not verifiable from the abstract | (c) AI and group ideation/decisions: — - **Caveats:** Abstract only. The design, group composition, measures and statistics are all unverified. The claim of "significant performance" cannot be checked, and the embodied-robot, gaze-focused contribution weighs more than the LLM facilitation. Weight low in synthesis. - **Screening check:** Provisionally meets criteria on the abstract. It is empirical (user study, n=50), involves multi-party groups of humans interacting with an LLM-powered robot that supports the discussion, and is in English. The method is reported only at abstract level, so inclusion is not fully verified; confirm group-based sessions and the reported method if full text becomes available. #### C64 — EngageCue: LLM cues via smart glasses for balanced group engagement - **Citation:** Liang, H., Liu, Y., Yang, J., Wang, X., Wong, F. Y. M., Zhao, K., & Fan, M. (2026). EngageCue: Enabling Balanced Engagement in Face-to-Face Group Collaboration via AI Cues Through Smart Glasses. International Journal of Human–Computer Interaction, 1–30. https://doi.org/10.1080/10447318.2026.2706908 - **Type:** journal — peer-reviewed: yes - **Full text read:** abstract only. Taylor & Francis is paywalled and there was no copy on the author's page or arXiv. Abstract from OpenAlex. - **Method:** Formative study (distilling three design goals), then a user study of the system. The comparison condition is not specified in the abstract. - **n:** 36 participants in 12 groups (abstract). - **Setting:** Face-to-face group collaboration. Smart glasses combine multimodal behaviour sensing, engagement-state assessment and LLM-based cue formulation to give real-time cues in a "sensing–analysis–feedback–interaction" loop (abstract). The LLM model is not verified. - **Findings:** 1. The system "significantly improved the balance of speaking engagement", but effects on overall engagement were "modest" (abstract). 2. Cue effectiveness depended on the timing and tone of cues (abstract). - **Relevance:** (a) — | (b) AI as a real-time participation-balancing facilitator in co-located groups | (c) more balanced speaking, which bears on inclusive group discussion (decision quality not reported) - **Caveats:** Abstract only, so statistics and design were not verified. Groups of 3. - **Screening check:** Meets criteria on the basis of the abstract (empirical, groups of humans with LLM-generated cues, n reported, English). #### OA86 — MUCA: Multi-User Chat Assistant - **Citation:** Mao, M., Ting, P., Xiang, Y., Xu, M., Chen, J., & Lin, J. (2024). Multi-User Chat Assistant (MUCA): a Framework Using LLMs to Facilitate Group Conversations. arXiv:2401.04883 (v4, 4 Oct 2024). https://doi.org/10.48550/arxiv.2401.04883 - **Type:** preprint — peer-reviewed: no (preprint) - **Full text read:** yes (https://arxiv.org/pdf/2401.04883, v4) - **Method:** System paper with case studies and a small user study; within-group comparison of a GPT-4 single-prompt baseline vs. the MUCA framework, counterbalanced order, two topics; descriptive metrics (engagement words, evenness, consensus rate) and participant ratings; no inferential statistics (§4.3). - **n:** 16 participants in 3 groups: Group-A and Group-B (4 people each, small-group comparison) and Group-C (8 people, MUCA-medium only); 1:1 female:male (§4.3.1, p. 7). - **Setting:** Text group chat; goal-oriented tasks ("indoor course set up in a community learning center", "interview agenda for hiring arts instructors"); chatbot based on GPT-4 that decides what to say, when to chime in and whom to address, including prompting lurkers and summarising (§3, §4.1). - **Findings:** 1. 9 of 16 participants (56.25%) felt the baseline bot chimed in excessively; no excessive chiming was reported for MUCA-small, though some said it chimed in too infrequently. MUCA-small "generally offers appropriate responses (13 out of 16)" (§4.3.2, p. 7). 2. Consensus rate (agreements over tasks): Topic A baseline 50% vs MUCA-small 66.7%; Topic B 50% vs 100%; MUCA-medium (8 people) 66.7%. Engagement words per conversation Topic A 426.5 vs 531.5 (Table 1, p. 8). 3. Overall user rating: MUCA-small exceeded baseline by 31.9% (Topic A: 69.44% vs 37.5%) and 11.1% (Topic B: 63.89% vs 52.78%) (§4.3.3, Table 1, pp. 7–8). - **Relevance:** (a) — | (b) LLM facilitating live multi-user text chats (timing, participation prompts, summaries) | (c) consensus rate and evenness of contribution, descriptive only - **Caveats:** Very small n (3 groups); consensus rates computed over a handful of tasks; "significantly higher" ratings stated without tests; comparison is LLM vs. LLM (no human-facilitator or no-bot control); preprint. - **Screening check:** Meets criteria (empirical, groups of 4–8 humans interacting with an LLM facilitator, method and n reported). #### C804 — KNIT-C: conversational AI facilitating group convergence - **Citation:** Wan, E., Mougenot, C., & Sadek, M. (2026). KNIT-C: Exploring Conversational AI Facilitation of Group Convergence Through Computational Boundary Objects. Proceedings of the 8th ACM Conference on Conversational User Interfaces (CUI '26), pp. 1–6. https://doi.org/10.1145/3816046.3816292 - **Type:** conference (short paper) — peer-reviewed: yes - **Full text read:** abstract only (ACM PDF blocked; no arXiv/author copy found) - **Method:** Exploratory evaluation of a conversational system in public workshop sessions (abstract); data collection and analysis details not verified. - **n:** 4 public workshop sessions, 14 non-expert participants (abstract). - **Setting:** KNIT-C elicits individual viewpoints, synthesises group perspectives into computational boundary objects, and guides synchronous multi-user dialogue toward negotiated outcomes (abstract); LLM not stated in abstract — not verified. - **Findings:** 1. Conversational engagement deepened individual reflection but created pacing and coordination challenges in group interaction (abstract). 2. AI-generated summaries anchored discussion, but their similarity required careful comparison rather than simplifying choice (abstract). 3. Participants often felt heard without fully endorsing outcomes — convergence as "productive disagreement" that keeps tension visible rather than consensus (abstract). - **Relevance:** (a) — | (b) conversational AI facilitating the convergence phase of group workshops | (c) AI summaries anchor group discussion; convergence without full endorsement - **Caveats:** Abstract only; small n (14); short paper; builds on the authors' earlier KNIT system (CHI '26, doi 10.1145/3772318.3791921) — possible overlap. - **Screening check:** Meets criteria on abstract — empirical, 4 workshop sessions with 14 participants interacting as groups with a conversational AI facilitator; method and n stated; English. #### C68 — KNIT: AI-generated boundary objects for convergence in interdisciplinary teams - **Citation:** Wan, E., Yin, C., Ito, A., Gao, Z., Jia, J., Taoka, Y., Saito, S., Sadek, M., & Mougenot, C. (2026). KNIT: Computational Boundary Objects for Real-Time Convergence in Interdisciplinary Teams. Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26), pp. 1–22. https://doi.org/10.1145/3772318.3791921 - **Type:** conference — peer-reviewed: yes - **Full text read:** abstract only. ACM DL returns 403 or a Cloudflare challenge and no arXiv or repository copy was found. Abstract from OpenAlex. - **Method:** Workshop-based evaluation with interaction-episode analysis using Carlile's 3T framework (syntactic, semantic, pragmatic boundaries) (abstract). - **n:** 7 early-stage healthtech teams, 28 participants, 190 interaction episodes analysed (abstract). - **Setting:** In-session workshops with interdisciplinary healthtech teams. KNIT externalises anonymised individual inputs into AI-generated shared artefacts (semantic clusters, stakeholder-centred problem reframings) for negotiation (abstract). The LLM model is not verified. - **Findings:** 1. KNIT supported knowledge boundary crossing at the syntactic (95.0%), semantic (86.3%) and pragmatic (84.8%) levels (abstract). The denominator (share of episodes) was not verified. 2. The authors conclude that "representational transformation rather than automation" is the main mechanism by which the AI enables convergence (abstract). - **Relevance:** (a) — | (b) AI-mediated workshop tool for anonymised input collection, clustering and reframing, i.e. core convergence facilitation steps | (c) supports negotiation and convergence across disciplinary perspectives - **Caveats:** Abstract only. No comparison condition is apparent. The percentages appear to be episode-coding shares, not outcome effects. - **Screening check:** Meets criteria on the basis of the abstract (empirical workshops, teams of about 4 humans with an AI mediator, n reported, English). #### OA2041 — The Empty Chair: LLM personas for missing perspectives in a student citizens' assembly - **Citation:** Fulay, S., Dimitrakopoulou, D., & Roy, D. (2025). The Empty Chair: Using AI Personas to Raise Missing Perspectives in Policy Deliberations. arXiv:2503.13812 (v2; NeurIPS 2025 Workshop PersonaLLM). https://doi.org/10.48550/arxiv.2503.13812 - **Type:** preprint / workshop paper — peer-reviewed: unclear (NeurIPS workshop; no formal proceedings verified) - **Full text read:** yes (https://arxiv.org/pdf/2503.13812, v2) - **Method:** Field case study with post-activity survey and pre/post attitude survey; qualitative participant comments; no control condition (p. 3–4, §5–6; limitations p. 6, App. A). - **n:** 19 undergraduate students in a 3-day campus citizens' assembly; tool used in a 1-hour session in 4 groups of 4–5 people, run by facilitators (p. 2–3, §3, §5). - **Setting:** In-person student assembly on university net-zero policy. Tool transcribed the discussion live and generated 3 absent-stakeholder personas plus points of disagreement and persona-voiced questions for an expert panel. The facilitator operated the tool; GPT-4o (p. 2–3, §4–5; p. 9, App. D). - **Findings:** 1. Post-activity ratings: engaging/thought-provoking m=5.83 (CI 5.44–6.23); useful m=5.61 (CI 5.19–6.03); sparked discussion of missing perspectives m=6.16 (CI 5.71–6.60); easy to use m=6.05 (CI 5.71–6.39). The scale range is not stated in the text read (p. 4, §6.1). 2. Pre/post: empathy toward people with different sustainability views +0.63 (6.11 vs 5.47); toward those who disagree +0.68 (5.53 vs 4.84); ability to hear and respect conflicting arguments +0.63 (6.00 vs 5.37). No significance tests reported. Per the Fig. 5 caption, pre = before day 1 and post = after day 1 of the assembly (p. 4, §6.2; p. 8, Fig. 5). 3. Concerns: personas may misrepresent groups or act as "a poor substitute for actually bringing in diverse voices", and some outputs were too generic or redundant. Facilitators advised framing the tool as a role-play aid or overlooked-perspective prompt rather than a stand-in for participants (p. 4, §6.1, §7). - **Relevance:** (a) facilitator use of AI: facilitators ran the tool live in breakout groups and gave feedback on how to frame it | (b) AI planning/running/supporting group sessions: real-time LLM support inside a deliberation session | (c) AI and group ideation/decisions: surfaced new perspectives and questions; self-reported perspective-taking rose - **Caveats:** Small student sample, no control group, no significance tests; the pre/post window covers a whole assembly day, so the change cannot be attributed to the tool; authors note these limits (App. A). - **Screening check:** PASSES (empirical; 4 human groups using an LLM in facilitated deliberation; method and n reported). #### OA2177 — Collective dialogues + AI to find common ground among Israeli and Palestinian peacebuilders - **Citation:** Konya, A., Thorburn, L., Almasri, W., Leshem, O. A., Procaccia, A. D., Schirch, L., & Bakker, M. A. (2025). Using Collective Dialogues and AI to Find Common Ground Between Israeli and Palestinian Peacebuilders. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT '25), pp. 312–333. https://doi.org/10.1145/3715275.3732022 - **Type:** conference — peer-reviewed: yes (FAccT). The authors describe it as a case study from a real-world project, "not academic research" (p. 2). - **Full text read:** yes (https://arxiv.org/pdf/2503.01769, arXiv v3; page numbers below refer to the arXiv PDF) - **Method:** Field case study of an iterative process (Apr–Jul 2024): online collective dialogues on Remesh, bridging-based ranking (max-min agreement plus a Community-Notes-style model), LLM distillation into collective statements, human expert and joint review, and a final vote with equal-power metrics. Micro-experiments with native-speaker blind preference on prompt language. No controlled evaluation (p. 4–7, §3; p. 8, §5). - **n:** Estimated lower bounds of unique engaged participants per dialogue cycle: 70, 28, 13 and 138 (3 uninational cycles plus 1 joint cycle); "around 100" in the uninational phase and "around 138" in the joint phase (p. 6, §3.3; p. 8, §4). Participants were civil-society peacebuilders: Israeli Jews, Palestinian citizens of Israel, and Palestinians from the West Bank and Gaza. - **Setting:** Run with the Alliance for Middle East Peace. Live Zoom plus Remesh dialogues, reopened asynchronously; Hebrew, Palestinian Arabic and English with Google Translate. LLMs: gpt-4-turbo-2024-04-09 (uninational phase) and gpt-4-0125-preview (joint phase) (p. 4, fn. 4; p. 6, §3.4). - **Findings:** 1. Joint phase: 5 demands to world leaders each had ≥90% agreement from every side (max-min agreement), and "An immediate ceasefire and release of all hostages" was ranked first on all sides. 5 messages to residents each had ≥84% agreement per side. The 10 public statements had 84–96% bridging support, 9 of them above 90% (p. 8, §4.2; §5). 2. In the language micro-experiments, native-speaker reviewers blind to the method preferred, "in all cases", statements generated with prompts and examples in their own language (p. 6, §3.4). 3. The authors say LLM distillation replaced multistakeholder deliberations "which can take months" with "an LLM-accelerated step of a few hours", compressing cycles to weeks. Participant feedback was mostly positive (anonymity allowed open sharing), but some said "Everything is so fast. It's hard to think" and wanted to see each other's faces (p. 9, §5). - **Relevance:** (a) facilitator use of AI: process facilitators and experts used LLMs to synthesise statements and reviewed them jointly | (b) AI planning/running/supporting group sessions: LLMs and translation supported multilingual collective dialogues in a real conflict setting | (c) AI and group ideation/decisions: AI-assisted common-ground statements reached high cross-group agreement - **Caveats:** No control or comparison; results "intentionally selective"; unique participants unknown (lower-bound estimates); the time saving is the authors' claim, not measured; Remesh-affiliated first author. - **Screening check:** PASSES (empirical field study; groups interacting via an LLM-supported deliberative process; method and n reported). Note: interaction is mediated and asynchronous-capable collective dialogue rather than face-to-face group discussion. #### C697 — Fine-tuned LLM (Chinchilla 70B) generating group consensus statements - **Citation:** Bakker, M. A., Chadwick, M. J., Sheahan, H. R., Tessler, M. H., Campbell-Gillingham, L., Balaguer, J., McAleese, N., Glaese, A., Aslanides, J., Botvinick, M. M., & Summerfield, C. (2022). Fine-tuning language models to find agreement among humans with diverse preferences. arXiv:2211.15006. https://doi.org/10.48550/arxiv.2211.15006 - **Type:** preprint — peer-reviewed: no (preprint; arXiv v1 read; any peer-reviewed version not verified) - **Full text read:** yes (https://arxiv.org/pdf/2211.15006, v1, 28 Nov 2022) - **Method:** Online experiments with small human groups: participants write opinions on UK political/moral questions, an LLM (supervised fine-tuned + reward model ranking under a utilitarian social-welfare function) generates candidate consensus statements, participants rate agreement and quality; comparisons against prompted zero-/few-shot LLMs, ablated fine-tuned baseline, and human opinions; mixed-effects logistic regression; opinion-exclusion analysis (§3–4). - **n:** 3,211 UK participants in 746 groups across training and evaluation (p. 4, §3.2); main baseline evaluation n = 530 (within-distribution, groups of five) and n = 267 (out-of-distribution, groups of four) (p. 7, §4.1); human-opinion comparison n = 189 (within-distribution, groups of four) (p. 8, §4.2); training data from 1,524 participants (p. 6). Reliability filter retained 1,028/1,524 (67%) training and 1,164/1,687 (69%) evaluation participants (p. 17, App.). - **Setting:** Crowdworkers (online platform, £15/h average), groups of 3–5, 5-minute opinion writing; 70B Chinchilla LLM fine-tuned by DeepMind; questions on UK current affairs (e.g., "Should we lower the speed limit on roads?") (pp. 2–4, 15–16). - **Findings:** 1. Participants preferred consensus statements from the fine-tuned SFT-Utilitarian model over all baselines, >70% over prompted LLMs (abstract, p. 1); agreement higher than the SFT-Base ablation (β = 0.12, SE = 0.047, z = 2.53, p = 0.011) (p. 8, §4.1). 2. On divisive rounds (~50% of rounds), 65.6% [61.9, 69.3] of model candidates were less divisive than the initial position statement, and in 40.8% [35.4, 46.2] a model candidate achieved unanimous (at least "somewhat agree") support (p. 8, §4.1). 3. Model statements were preferred over the mean human opinion 78% [75%, 80%] of the time, and the best model candidate over the best-rated human opinion 65% [61%, 69%] (p. 8, §4.2); participants whose opinions were silently excluded agreed less (Likert difference 0.47, CI [0.21, 0.73]) while quality ratings were unaffected (p. 9, §4.3). - **Relevance:** (a) — | (b) LLM performing a core facilitation function (drafting consensus statements from a group's opinions) | (c) LLM-mediated consensus statements increased group agreement on divisive questions; exclusion of voices reduces their agreement - **Caveats:** Groups do not deliberate directly — interaction is asynchronous via written opinions and ratings; crowdworkers, UK only, groups of ≤5 (p. 10, §5.1); preprint version read; DeepMind-internal model. - **Screening check:** Meets criteria — empirical, groups of 3–5 humans whose opinions an LLM aggregates into consensus statements they rate (group decision/consensus function); method and n reported; English; 2022. #### OA1997 — GPT-4o for moderating large-scale online deliberation (SAC 2025) - **Citation:** Babatunde, I. D., Nnanna, O. M., & Klein, M. (2025). Moderating large scale online deliberative processes with large language models (LLMs): Enhancing collective decision-making. In Proceedings of the 40th ACM/SIGAPP Symposium on Applied Computing (SAC '25), Catania, pp. 996–1003. ACM. https://doi.org/10.1145/3672608.3707925 - **Type:** conference — peer-reviewed: yes (ACM SAC) - **Full text read:** yes (MIT DSpace copy, https://hdl.handle.net/1721.1/162200, via dspace.mit.edu/server/api/core/bitstreams/af2868db-8cad-4307-a0ba-94b25f28e147/content) - **Method:** offline evaluation on a corpus: GPT-4o (temperature 0) performed four moderation subtasks — answer mining/unbundling, answer classification (solution/metric/barrier), argument mining, argument PRO/CON classification — under zero-shot, two few-shot variants (4 examples each) and few-shot chain-of-thought prompting; outputs compared with consensus annotations by three trained expert annotators using precision, recall and F1 (accuracy excluded because of class imbalance) (§3, pp. 998–1001). - **n:** 72 purposively selected test cases (not random) drawn from a dataset of "over 500 conversation scripts" (§3.1, p. 998). - **Setting:** transcripts from collaborative workshops with volunteer facilitators and participants contributing to a "deliberation map" on the Deliberatorium platform (topics such as cross-departmental collaboration, open-access research, professional growth) (§3.1, p. 998). - **Findings:** 1. Zero-shot: argument mining precision 74%, recall 64%, F1 69%; answer mining recall 86% but precision 57% (arguments misclassified as answers); answer classification recall 87%, F1 68%; argument classification F1 68% (§4.1.1, p. 1001). 2. Few-shot prompting improved answer-related tasks (Few-shot 1: answer-mining recall 91%, answer-classification F1 77%; Few-shot 2: answer-mining precision 74%, answer-classification recall 88%, F1 73%), but argument classification stayed weak (Few-shot 1 precision 49%, F1 52%; Few-shot 2 precision 62%, F1 67%) (§4.1.2–4.1.3, pp. 1001–1002). 3. Few-shot chain-of-thought gave the lowest argument-classification scores (precision 35%, F1 41%) despite high recall on answer tasks (87% answer mining, 83% answer classification); authors conclude GPT-4o is well suited to unbundling/categorising ideas but limited on "subtle argument structures" (§4.1.4, §5.1, p. 1002). - **Relevance:** (a) — | (b) LLM performing facilitator/moderator structuring functions (idea unbundling, categorisation, argument mapping) on real workshop deliberation data | (c) indirect: structuring contributions for collective decision-making - **Caveats:** small, non-random test set (72); single model; numbers taken from text because Table 1 is an image in the extracted PDF; some internal inconsistency (text attributes the 74% argument-mining precision to both zero-shot, §4.1.1, and Few-shot 2, §5.1; "best" argument-classification precision stated as zero-shot 73%, §4.2); no human-in-the-loop or live deployment; no inter-annotator agreement statistic reported. - **Screening check:** meets criteria — offline evaluation of an LLM performing a facilitation/moderation function on real group-deliberation data, n (test cases) reported. #### OA1999 — Opportunities and risks of LLMs for scalable deliberation with Polis - **Citation:** Small, C. T., Vendrov, I., Durmus, E., Homaei, H., Barry, E., Cornebise, J., Suzman, T., Ganguli, D., & Megill, C. (2023). Opportunities and risks of LLMs for scalable deliberation with Polis. arXiv:2306.11932. https://doi.org/10.48550/arXiv.2306.11932 - **Type:** preprint (technical report, The Computational Democracy Project & Anthropic) — peer-reviewed: no (preprint) - **Full text read:** yes (https://arxiv.org/pdf/2306.11932, v1, 20 Jun 2023) - **Method:** exploratory pilot experiments / offline evaluation on existing deliberation data: Claude prompted for topic modelling, summarisation, consensus-statement generation, vote prediction, LLM "voting" to measure ideological bias, and large-context (100K-token) summarisation; outputs judged qualitatively against the authors' manual analysis, plus a calibration analysis and a bias t-test; authors state the purpose is "more qualitative than quantitative" (§2, p. 9; Appendix A, p. 43). - **n:** one main dataset: the 2018 Bowling Green, Kentucky Polis conversation with ~2,000 of 65,000 residents participating (§2, p. 9); number of comments/votes analysed per experiment not reported as a single figure (topic modelling in batches of 50 comments, Appendix A.1). - **Setting:** Polis online deliberation platform; Claude models from the public API as of 15 Mar 2023 (8K-context research model for topic modelling) and the 100K-context model as of 12 Jun 2023 (Appendix A, p. 43; §2.1 footnote, p. 9). - **Findings:** 1. Topic modelling and summarisation outputs "matched our expectations from manual analysis" / "concur with manual analysis", but summaries sometimes misstated which side a group took (e.g., on the opioid crisis); with a 100K context window, summaries were "more detailed and nuanced" than batch-merged ones — context limits materially affected quality (§2.1.1 p. 10; §2.2.1 p. 12; §2.5 p. 21). 2. Vote prediction: an LLM given a participant's voting history was "nearly perfectly statistically calibrated" at predicting agreement, and in a majority of cases ≥90% confident; the authors "reject as invalid on both ethical and performance grounds" replacing participants with simulations (§2.4.1–2.4.2, pp. 17–18). 3. Bias: when Claude voted on the comments, its votes were significantly closer to the more socially progressive opinion group (Group 0) (t-test p = 1.86 × 10⁻⁴), which the authors flag as problematic for a "multi-partial" facilitation agent (§3.2, pp. 23–24). LLM-generated consensus statements could not be validated because they were not tested in a live conversation (§2.3.1, p. 15). - **Relevance:** (a) — | (b) LLM performing facilitator/reporting functions (topic mapping, summarising, consensus statements, moderation) for large-scale deliberation | (c) risks for group decision outputs: summarisation errors, political bias, temptation to simulate participants - **Caveats:** preprint by platform and model developers (Anthropic co-authors evaluating Claude); mostly qualitative judgments by authors; single dataset; no live deployment or human-participant outcome measures; models now outdated. - **Screening check:** meets criteria — offline evaluation of an LLM performing facilitation functions on real group-deliberation data; method described, data source and scale reported (though per-task n is not precise). #### OA785 — Wizard-of-Oz "GenAI-style" conversational cues in triad brainstorming - **Citation:** Rayan, J., Kanetkar, D., Gong, Y., Yang, Y., Palani, S., Xia, H., & Dow, S. P. (2024). Exploring the Potential for Generative AI-based Conversational Cues for Real-Time Collaborative Ideation. Proceedings of the 16th Conference on Creativity & Cognition (C&C '24), pp. 117–131. https://doi.org/10.1145/3635636.3656184 - **Type:** conference — peer-reviewed: yes (ACM C&C) - **Full text read:** abstract only (ACM DL listed as hybrid OA but PDF/fullHtml blocked by Cloudflare/403; abstract via OpenAlex) - **Method:** Design probe study with Wizard-of-Oz: a human "wizard facilitator" playing the role of an AI agent oversaw remote triad brainstorming and sent real-time text cues into the ideators' interface; thematic analysis of conversations, cues and facilitators' post-hoc reflections (abstract). - **n:** 42 people in 14 triads (abstract). Number of wizard facilitators not verified. - **Setting:** Remote meeting brainstorming via a custom design probe; no actual LLM generated the cues (humans simulated the AI) (abstract). - **Findings:** 1. 44% of cues sent by facilitators were dismissed or ignored because ideators did not notice the cue update (abstract). 2. When cues were noticed, certain facilitator strategies impacted the conversation more than others (abstract; which strategies — not verified). 3. Authors derive design opportunities for GenAI systems supporting real-time creative collaboration (abstract). - **Relevance:** (a) facilitator use of AI: — (humans act as the AI) | (b) AI planning/running/supporting group sessions: directly probes an AI-facilitator role sending real-time cues in group ideation | (c) AI and group ideation/decisions: noticing/uptake of facilitation cues; outcome effects not verified. - **Caveats:** Wizard-of-Oz — no generative AI actually used; only abstract read; page-level details not verified. - **Screening check:** Borderline: meets group criterion (triads) and addresses AI facilitation, but no generative AI/LLM was actually involved (human wizard simulating AI). Include only if WoZ simulations of LLM facilitation are admissible; otherwise FAILS: no actual generative AI/LLM involvement (Wizard-of-Oz). #### OA149 — Facilitating data conversations with generative AI - **Citation:** Diederich, M., & Nguyen, H. (2024). Facilitating Data Conversations with Generative AI. In Proceedings of the 18th International Conference of the Learning Sciences (ICLS 2024), pp. 2125–2126. https://doi.org/10.22318/icls2024.307157 - **Type:** conference (2-page poster/short paper) — peer-reviewed: yes (ISLS proceedings; review type not verified) - **Full text read:** yes (https://repository.isls.org//bitstream/1/10896/1/ICLS2024_2125-2126.pdf) - **Method:** Within-dyad comparison, Wizard-of-Oz: each dyad held two 15-min data conversations in random order, one with four predetermined (ChatGPT-generated, author-revised) questions and one with real-time AI facilitation (researchers pasted live transcripts into ChatGPT and relayed its questions at pauses or every four minutes); qualitative/deductive coding of utterances; descriptive statistics only (Methods, p. 2125). - **n:** 4 dyads (8 graduate students in Instructional Design: five Master's, three PhD), 120 min of audio (p. 2125). - **Setting:** Microsoft Teams conversations about recreation decisions and their environmental impacts, targeting data literacy and critical data literacy (Data Feminism); ChatGPT (version not stated) (p. 2125). - **Findings:** 1. 79% of ChatGPT's real-time questions built on the conversation by quoting or referencing what participants had said; all reflected the prompt's focus on data practices, but only 50% showed critical data literacy and 36% focused on recreational/environmental topics (Findings, pp. 2125–2126). 2. With AI facilitation, participants mentioned data variability 5 times (predetermined: 0) and data communication 5 times (predetermined: 1); authors report no statistical tests due to small sample (p. 2126). 3. Critical data literacy ideas (e.g., ethics/privacy 11 utterances, labour behind data 8, emotion 7) occurred with AI facilitation but less often than data practices (p. 2126). - **Relevance:** (a) — | (b) LLM generating real-time facilitation questions for small-group (dyad) discussions | (c) content of group talk, descriptive only - **Caveats:** Wizard-of-Oz; 4 dyads; student sample; 2-page proceedings paper; educational context. - **Screening check:** Meets criteria (empirical; dyads of two humans in a facilitated discussion with LLM-generated facilitation; method and n reported). Borderline on the education exclusion, but AI facilitates a structured group discussion and group-level talk is analysed, so it fits the education inclusion rule. #### C718 — Read the Room: timing of proactive LLM agent participation in group chats - **Citation:** Yoo, S. Y. (2026). Read the Room: Exploring When LLM Agents Should Participate in Everyday Mixed-Context Group Chats on Mobile Messaging Platforms. Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA '26), pp. 1–6. https://doi.org/10.1145/3772363.3798392 - **Type:** conference (extended abstract) — peer-reviewed: yes (juried; extended abstract) - **Full text read:** abstract only (ACM PDF blocked; no arXiv/author copy found) - **Method:** Exploratory user sessions in a ChatGPT group chat where the agent proactively participated based on context; qualitative perception data (abstract). Details not verified. - **n:** 6 dyads of acquaintances (12 participants implied) (abstract). - **Setting:** Each pair engaged in casual conversation and a travel-planning task in a ChatGPT group chat on mobile messaging (abstract). - **Findings:** 1. Proactive agent participation was perceived as appropriate when the conversation lacked a clear next step and when the information need was already aligned (abstract). 2. It was perceived as disruptive during active opinion exchange, in both contexts but especially in casual talk (abstract). - **Relevance:** (a) — | (b) when an AI should intervene in multi-person conversation (intervention timing, relevant to AI facilitation) | (c) — - **Caveats:** Abstract only; very small sample (6 dyads); dyads of humans plus agent; everyday chat, not a facilitated session. - **Screening check:** Borderline pass — empirical, method and n stated, English; 2 humans interacting as a group with an LLM agent (human pairs, not a human–AI dyad). Setting is casual/planning group chat rather than a facilitated work session; include only if casual group chats count as "group interaction" under criteria. #### S2-198 — AgoraAI: voice-to-voice multi-persona LLM framework for multi-human meetings - **Citation:** Concha-Sánchez, A., Bernal-Millan, J. A., Hernández-Muñiz, A., & Gadi, S. K. (2026). AgoraAI: An Open-Source Voice-to-Voice Framework for Multi-Persona and Multi-Human Interaction. Applied Sciences, 16(4), 2120. https://doi.org/10.3390/app16042120 (published 21 February 2026; author surnames per Crossref and the article's contribution initials "A.C.-S.") - **Type:** journal — peer-reviewed: yes - **Full text read:** yes (MDPI HTML https://www.mdpi.com/2076-3417/16/4/2120/htm via reader-proxy markdown, pdf3/S2-198.md; cited by section number. Inline math did not render, so some numbers such as the explicit "n = 15" symbol and the §3.5 latency mean/SD are missing from the proxy text; the total of 15 is confirmed in §8 and by the cohort breakdown in §5.1) - **Method:** System/engineering paper plus (i) two qualitative field deployments reported as case studies from transcripts and observations (§4.5, §4.6) and (ii) an exploratory mixed-methods pilot survey after sessions (10-point Likert scale plus open-ended questions; descriptive statistics only) (§5). Four further scenarios (§4.1–4.4) are described only anecdotally. Ethics review was waived (IRB statement). - **n:** Pilot survey: 15 users (§8). Academic cohort = 3 graduate students + 5 professors (8). Professional cohort = 3 staff + 4 managers (7) (§5.1). Deployment 1 (thesis seminar): 5 master's students presented one after another to a hybrid panel of 2 AI personas + 2 human professors, with a human operator handling turn-taking (§4.5.1–4.5.2). Deployment 2 (corporate coordination meeting): 2 human department heads (Head of Design, Head of Production) + 2 AI agents (Project Manager, Secretary), time-boxed to 20 min (§4.6.1–4.6.2). - **Setting:** Mexico: master's program in Mechanical Engineering at Universidad Autónoma de Coahuila, and Industria Sigrama S.A. de C.V. (Torreón). The seminar ran in Spanish. A single Google Gemini 2.0 Flash backend handles STT and LLM, switching personas by context injection. Node.js/Socket.IO orchestration uses a "dual-queue" design. A human operator manages turn-taking (§3.3, §6.2). - **Findings:** 1. In the corporate meeting, the AI Project Manager opened with the constraints, asked the two human heads for commitment dates and negotiated milestones (Design Freeze by 18 Dec, contingent on approvals by 15 Dec; purchase orders by 20 Dec). The AI Secretary read the agreements back for validation. The authors report that the agents "successfully managed the meeting agenda" and finalized the schedule within the 20-min time box. This is the authors' qualitative claim, with no outcome measure (§4.6.2–4.6.3). 2. Pilot survey (1–10 scale): 75% of academic users said AgoraAI helped them understand the subject better. 100% of professional users said it made meetings more efficient. 75% of academic users would use it again. Professional recommendation score was 9.4/10. Interface intuitiveness 8.0, ease of following conversations 8.5, clarity of AI responses 8.1–8.8, AI explaining concepts 9.4, usefulness of transcripts/summaries 8.5–9.0 (§5.2, Table 4; per-group item breakdown not verified because Table 4 did not render). 3. Constraints: the system needs a human operator for turn-taking. One operator can handle at most 4 concurrent agents and 2 human speakers, which limits use to facilitated rather than unmoderated sessions (§6.2). End-to-end latency is about 6.9 s and causes "turn-taking rigidity", making the system unsuitable for rapid, interruptible conversation (§6.3). No hallucinations were identified in the pilot (§6.3). - **Relevance:** (a) facilitator use of AI: — (a human operator runs the system; not studied as a facilitator practice) | (b) AI planning/running/supporting group sessions: LLM personas act as meeting facilitator/project manager, secretary and panelists in two live multi-human sessions, with a usability/acceptance survey | (c) AI and group ideation/decisions: qualitative account of an AI-facilitated scheduling negotiation reaching agreed commitment dates; no measured decision quality - **Caveats:** Very small convenience sample (15) with descriptive statistics only and no comparison condition. Case-study outcomes are author-reported. Three of the four authors were employed by Industria Sigrama, the deployment company, which also funded the APC (conflict of interest). The human operator is a Wizard-like element for turn-taking. The proxy text lost inline math and table contents. - **Screening check:** Meets criteria. The paper is empirical, and ≥2 humans interact as a group with generative-AI (Gemini) agents in live sessions (2 department heads + AI facilitator; students + professors + AI panelists). It reports its methods and n (pilot n=15; two deployments) and is in English. The evidence is weak: an exploratory pilot plus case studies. #### C779 — AI-assisted (RAG multi-agent) revision of Delphi statements vs. human revision - **Citation:** Nello, E., Tedone, F., Caproni, E., Cafiero, D., Manellari, S., & Rocco, P. (2026). A Controlled Comparison of Human and AI-Assisted Automated Revision of Delphi Statements on RNA-Based Medicines: Parallel, 2-Arm Study. JMIR Medical Informatics, 14, e90228. https://doi.org/10.2196/90228 - **Type:** journal — peer-reviewed: yes - **Full text read:** yes (PMC HTML, https://pmc.ncbi.nlm.nih.gov/articles/PMC13408467/; no page numbers — sections cited) - **Method:** Parallel 2-arm Delphi study; panelists stratified and randomized (blind to allocation) to human revision (arm A) vs. AI-assisted revision (arm B) of statements below a 75% agreement threshold after round 1; descriptive comparison of pass rates, no inferential tests (Methods). - **n:** 50 international panelists (clinicians, researchers, patient representatives), 25 per arm; 28 statements; 2 human reviewers (non-panelists) supervising revisions (Abstract; Methods). - **Setting:** Online Delphi on RNA-based medicines (5-point Likert, 2-week round); AI pipeline with ReferenceDetector, Summarizer, hybrid dense/BM25 RAG with cross-encoder reranking and Refiner agent producing revised statements, change logs and citations; human-in-the-loop approval. The specific LLM is not named in the text read ("a specific LLM", Limitations) — not verified. - **Findings:** 1. Round 1 consensus: arm A 71.4% (20/28) vs. arm B 46.4% (13/28) (mean Likert 4.24 vs 3.90); after one revision round, arm A 92.9% (26/28, 95% CI 76.5–99.1%) vs. AI arm 85.7% (24/28, 95% CI 67.3–96.0%); gain +21.4 vs +39.3 percentage points; final gap 7.2 points (Results; Abstract). 2. None of the AI-revised statements was rejected by reviewers as inaccurate; only 4 required minor rewording; the reference-request trigger fired once, correctly, with no false positives or negatives (Results). 3. Both arms failed to rescue 2–4 statements with substantive disagreement (2/28 human arm, 4/28 AI arm remaining below threshold); authors attribute the larger AI-arm gain partly to a more critical AI-arm panel at baseline; time/cost savings were not measured (Results; Discussion; Limitations). - **Relevance:** (a) — | (b) LLM pipeline performing the facilitator/moderator step of a Delphi (synthesising panel comments, revising statements) | (c) AI-revised statements reached near-equivalent consensus to human-revised ones - **Caveats:** Descriptive only, no significance tests; baseline imbalance between arms; single domain; model unnamed; human reviewers edited/approved AI output; efficiency not measured. Delphi panel is an asynchronous, anonymous group (no direct interaction). - **Screening check:** Meets criteria — empirical randomized 2-arm study; 50 experts form Delphi panels (structured group consensus process) in which an LLM pipeline performs the facilitator's revision function; method and n reported; English. ## 2. Facilitators and organisers using generative AI in real workshops, sprints, assemblies and debriefs #### GS1 — Human-GenAI Value Loop in design sprints - **Citation:** Grange, C., Demazure, T., Ringeval, M., Bourdeau, S., & Martineau, C. (2025). The Human-GenAI Value Loop in Human-Centered Innovation: Beyond the Magical Narrative. Information Systems Journal, 36(1), 29–51. https://doi.org/10.1111/isj.12602 - **Type:** journal — peer-reviewed: yes (text read is the arXiv preprint version 2407.17495; page numbers below refer to that preprint, not the ISJ version) - **Full text read:** yes (https://arxiv.org/pdf/2407.17495) - **Method:** Collaborative practice research (practitioner-facilitator + four academic researchers); multi-case field study of three GenAI-augmented design sprints (April–November 2023) documented by the facilitator (activities, timing, prompts/outputs, participant behaviour, outcomes), plus semi-structured interviews (~1 h each); interpretive analysis in twelve sensemaking sessions (§3.1, pp. 7–9). - **n:** 3 design sprints in 3 organisations; 5 post-sprint interviewees (p. 8). Project #1 team: 7 human participants plus ChatGPT as "J.A.K.E." (p. 12); Project #2: 8 participants voted on artifacts (p. 16). Project #3 team size not verified. - **Setting:** Real innovation design sprints (creativity school; municipality employee onboarding; public TV company digital experience for young users) run by one facilitator using ChatGPT, later custom GPTs and DALL-E, in three-hour workshops and between sessions (§3.3, pp. 11–19). - **Findings:** 1. In Project #1 ChatGPT acted as an anthropomorphised team member (prompted live by a human co-facilitator); its proposed problem statement received 75% of the team's votes in the alignment workshop (p. 13). The facilitator judged real-time prompting a "headache" and the anthropomorphised-participant approach questionable; GenAI was repositioned as a facilitator support tool (pp. 14–15). 2. In Project #2, with a prepared prompt repertoire, the facilitator used ChatGPT to seed ideas, rearticulate/classify human ideas and synthesise concepts; seven out of eight participants voted in favour of GenAI-generated user-test artifacts (pp. 15–16). Managing data and prompts in parallel with running the workshop remained challenging (p. 16). 3. In Project #3 the facilitator used task-specific custom GPTs (e.g., theme extraction for affinity clustering), a dynamic "project context document" as collective memory, DALL-E visuals, and tagged all content by origin (untouched / reformulated by AI / generated by AI); an AI concept resembling a human one was withheld to avoid voting bias; a team member reported transparency sustained trust and rejected AI deciding pivotal votes (pp. 17–19). Authors conclude human input matters most in divergent tasks, GenAI adds most value in convergent tasks (synthesis, alignment) (pp. 3–4). - **Relevance:** (a) facilitator use of AI: detailed account of one facilitator's evolving GenAI practice across sprints | (b) AI planning/running/supporting group sessions: GenAI as participant vs facilitator tool in live workshops | (c) AI and group ideation/decisions: voting on AI-generated statements/artifacts, divergent vs convergent roles - **Caveats:** Single facilitator (practitioner co-author), 3 cases, only 5 interviews; qualitative, no comparison condition; practitioner-researcher role may bias interpretation; vote shares are single-instance observations. - **Screening check:** Meets criteria — empirical field case study with stated method and sample (3 sprints, 5 interviews), multi-person teams in facilitated workshops, generative AI (ChatGPT, GPTs, DALL-E), English. #### OA379 — How professionals integrate GenAI into team meetings - **Citation:** Schätzle, A., Abdo, M., Gräf, M., & Buxmann, P. (2026). Understanding How Professionals Integrate Generative AI into Team Meetings: A Qualitative Study. Proceedings of the 59th Hawaii International Conference on System Sciences (HICSS), pp. 444–453. https://doi.org/10.24251/hicss.2026.053 (https://hdl.handle.net/10125/111445) - **Type:** conference — peer-reviewed: yes - **Full text read:** yes (ScholarSpace PDF, https://hdl.handle.net/10125/111445) - **Method:** qualitative; five semi-structured focus groups (avg. 35 min, via MS Teams, early 2025), guide sensitised by Allen et al. (2016) meeting model; Mayring qualitative content analysis, two coders (§3.1, p. 447) - **n:** 20 professionals in 5 focus groups (E1–E7, E8–E10, E11–E14, E15–E17, E18–E20) (§3.1 p. 447; §3.2 p. 448) - **Setting:** Germany-based professionals from consulting, software development and business operations with hands-on GenAI use in real (virtual/hybrid) team meetings; tools mentioned include Microsoft Copilot and ChatGPT (§3.2 p. 448; §4.3 p. 449) - **Findings:** 1. Offloading note-taking/summarising to GenAI was reported to raise meeting satisfaction and engagement and to let juniors contribute instead of being confined to minute-taking ("flattening of hierarchical structures"); traditional roles such as note-taker or moderator described as giving way to more fluid role distribution (§4.1, pp. 448–449). 2. GenAI gave real-time knowledge support (past notes, on-the-spot summaries), but experts reported interruptions of conversational flow and uneven use/trust across members, creating fragmented knowledge access ("Some have information from GenAI and others do not") (§4.2, p. 449). 3. Common uses: automated summaries, interim decision capture, onboarding support, often with Copilot; outputs must be checked for accuracy in high-stakes meetings; satisfaction depended on output quality and poor outputs caused frustration (§4.1, §4.3, p. 449). - **Relevance:** (a) practitioners' (incl. moderators/leads) use of GenAI in meetings | (b) GenAI supporting live team meetings (summaries, decision capture) | (c) qualitative: participation/role effects, no decision-quality data - **Caveats:** self-report, purposive sample of experienced users; findings reported at group-ID level, no counts; single country. - **Screening check:** Meets criteria — empirical qualitative study, n=20 reported, meeting professionals as respondents about GenAI use in team meetings, English. #### OA356 — Customised LLM meeting assistants in Agile (Austrian Post) - **Citation:** Cabrero-Daniel, B., Herda, T., Pichler, V., & Eder, M. (2024). Exploring Human-AI Collaboration in Agile: Customised LLM Meeting Assistants. In Agile Processes in Software Engineering and Extreme Programming (XP 2024), Lecture Notes in Business Information Processing, pp. 163–178. https://doi.org/10.1007/978-3-031-61154-4_11 - **Type:** conference (LNBIP) — peer-reviewed: yes - **Full text read:** yes (arXiv:2404.14871v1, https://arxiv.org/pdf/2404.14871; page numbers = preprint pages) - **Method:** Action research in industry: iterative design and in-situ interventions of two GPT-4 assistants (Agile Release Train Coach assistant for PRIME planning/refinement meetings; Scrum Team Assistant Tool giving live recommendations in Daily Scrums), with observation notes, a pre-intervention survey and anonymous post-meeting surveys (§3, §4). - **n:** Pre-intervention survey: 39 responses from Scrum teams; interventions with 3 development teams (Daily Scrum: four design iterations with three teams) and three PRIME iterations; one post-meeting survey reported with 7 respondents (Team 1) (§3.3 p. 6; §4 pp. 8–10; Tables 2–3). - **Setting:** Austrian Post Group IT, Digital Logistics Platform; online meetings (Microsoft Teams); Azure OpenAI Studio GPT-4 (§3.1–3.2). - **Findings:** 1. First Daily Scrum intervention: messages in the Teams chat were "overwhelming and rather distracting"; 2 of 7 respondents felt annoyed, but 4 of 7 liked being warned about overly detailed discussion and a predicted overrun of the 15-minute time-box (§4.2, p. 10). 2. Rewording "warnings" as shorter "recommendations" and using disappearing pop-ups changed tone ("friendlier, tactful") and later interventions were well received; participants valued an end-of-meeting summary, yet some felt "observed" or that "something inhuman" forced a pattern on them, and one said the Scrum Master should decide which AI recommendations are shared (§4.2, p. 10). 3. In the first PRIME intervention a third of the generated insights were wrong because of an outdated input value (human error); a participant still called the AI "very handy"; 76% of PRIME participants reported feeling excited or curious about AI assistants (§4.1, pp. 7–8). - **Relevance:** (a) Agile coaches/Scrum teams using LLM assistants to prepare and run meetings | (b) LLM live meeting assistance (time-boxing, staying on topic, summaries) in real team meetings | (c) — - **Caveats:** Single organisation; small, partly unreported respondent numbers per intervention; practitioner co-authors (action research); outcomes mostly qualitative. - **Screening check:** Meets criteria (empirical field study with real multi-person meetings using GPT-4 assistants; method and n reported, though per-intervention n is partial). #### C217 — LLM use in Scrum management (Brazilian practitioner survey) - **Citation:** Perkusich, M., Albuquerque, D. W., Araújo, A. A., Paixão, M., Gheyi, R., Kalinowski, M., & Perkusich, A. (2026). Adoption of Large Language Models in Scrum Management: Insights from Brazilian Practitioners. Lecture Notes in Business Information Processing (Springer), pp. 255–273. https://doi.org/10.1007/978-3-032-22375-3_16 (arXiv:2602.08192) - **Type:** conference (LNBIP proceedings) — peer-reviewed: yes. The full text read is the arXiv version. - **Full text read:** yes (https://arxiv.org/pdf/2602.08192) - **Method:** Cross-sectional online survey (Google Forms, October–November 2025) with non-probabilistic convenience and snowball sampling, descriptive statistics (p. 5–6 of PDF, §3). - **n:** 70 respondents. 49 actively use Scrum and 33 reported using LLM assistants in Scrum practice (abstract, p. 1 of PDF). Of respondents, 47% developers, 12% Scrum Masters, 8% QA/testers and 6% Product Owners (p. 7 of PDF). - **Setting:** Brazilian software professionals. Tools named: ChatGPT (100%), Gemini (82%), Copilot Chat (67%), Claude and DeepSeek (27%) (p. 7 of PDF). - **Findings:** 1. For Scrum events (the facilitated team meetings), LLM use is led by summarising what happened during events (68% use or plan, 36% already using) and synthesising Sprint Review feedback (64%, 42% using). Designing agendas and timeboxes (54%) and identifying impediments (48%) are adopted more cautiously (p. 9 of PDF, §4 RQ2). 2. Perceived benefits: productivity and quality 78% each, reduced effort on repetitive tasks 75%. Collaboration benefits (64%) and transparency (57%) are more modest. Scrum Masters rate AI useful less often (56%) than developers (84%) (p. 10–11 of PDF). 3. Risks: "almost correct" outputs 81%, confidentiality 63%, hallucinations 59%. About 30% worry about eroded Scrum understanding and 26% about reduced human connection (p. 11 of PDF). - **Relevance:** (a) practitioners, including Scrum Masters, use LLMs to prepare agendas and timeboxes and to summarise events | (b) LLM support for preparing and summarising team ceremonies (sprint review, retrospective-type events) | (c) — - **Caveats:** Small convenience sample from one country. Percentages appear to be based on the 33 LLM users (base not always stated). Only 12% are Scrum Masters. Self-report only. - **Screening check:** Borderline pass. Respondents are practitioners who run facilitated team events (Scrum ceremonies) and report generative-AI use for event preparation and synthesis; method and n are reported; English. Individual rather than group use of AI, so it counts under the "facilitators/meeting professionals as respondents" criterion only if Scrum practitioners are accepted as such. #### OA1540 — ChatGPT in MICE event planning and marketing content (ICOI 2025) - **Citation:** Chang, W.-Y. (2025). Unlocking the power of generative AI: Applications of ChatGPT in event planning and marketing content creation. Proceedings of Forum Manajemen Indonesia International Conferences (International Conference of Organizational Innovation 2025, FMI Vol. 2), pp. 546–560. https://doi.org/10.47747/fmiic.vi2.2973 - **Type:** conference — peer-reviewed: unclear - **Full text read:** yes (https://www.journal.jis-institute.org/index.php/fmiic/article/download/2973/2057) - **Method:** qualitative: semi-structured Zoom interviews (45–60 min) with a pre-survey, plus two document-based case studies of 2023 events; thematic analysis (Braun & Clarke) in NVivo, two researchers independently coded a subset (§3, pp. 549–551). - **n:** 5 MICE professionals (event planners, marketing managers, content strategists) recruited by purposive sampling; 2 event cases (ASEAN Digital Transformation Dialogue 2023, Jakarta; Mekong Eco-Creative Festival 2023, Chiang Rai) (§3, pp. 549–550). - **Setting:** Southeast Asian MICE industry (convention bureaus, event consultancies, marketing agencies); ChatGPT or comparable generative AI used for guidebooks, speeches, press releases, programme descriptions, translation, brainstorming (§3, p. 550). - **Findings:** 1. Time savings: for the ASEAN Dialogue, organisers "reported saving up to 40% of their usual time during the production of event materials"; AI was viewed as able to "accelerate the first 70% of content creation", with a mandatory human fact-check/"AI review protocol" in both cases because of errors such as misattributed affiliations, outdated facts and misquoted speaker titles (§4.1, p. 551; §4.3, p. 553). 2. Ideation: ChatGPT "was widely adopted to overcome idea stagnation, especially in brainstorming meetings" (themes, slogans, storytelling), described by one interviewee as "brainstorming with a limitless partner"; junior staff contributed ideas by prompting ChatGPT and presenting options to the group (§4.1, p. 552). 3. Team dynamics: participants reported more horizontal content development and increased psychological safety for less experienced staff, and prompt development was done as a team exercise during a brainstorming session at the ASEAN Dialogue (§4.4, pp. 554–555). - **Relevance:** (a) event/meeting professionals' self-reported use of ChatGPT for planning content and in brainstorming meetings | (b) — (agenda design appears only in the abstract and literature review; no agenda-specific findings in the results) | (c) self-reported easing of creative blocks and broader participation in team ideation - **Caveats:** n = 5; no quotes attributed to individual participants; percentages ("up to 40%", "first 70%") are self-reports without measurement detail; "comparison with similar past events" mentioned in method but no comparative data reported; focus is mostly marketing content, not facilitation; peer-review status unclear. - **Screening check:** meets criteria marginally — empirical interviews with meeting/event professionals as respondents, method and n reported; relevance to facilitation proper is limited (event content production more than session facilitation). #### GS9 — Generative AI reports for simulation team debriefings - **Citation:** Tscholl, D. W., Ebensperger, M., Rahrisch, A., Wang, H., Heckel, H., Thomasius, M., Kaserer, A., Grande, B., Seelandt, J. C., & Kolbe, M. (2026). Generative AI in simulation debriefings: an exploratory study using the Team-FIRST framework and qualitative feedback from simulation experts and learners. Advances in Simulation, 11, 14. https://doi.org/10.1186/s41077-026-00407-0 - **Type:** journal — peer-reviewed: yes - **Full text read:** yes (Europe PMC full-text XML of PMC12924402; no page numbers in this format, sections cited) - **Method:** Qualitative exploratory study; thematic analysis of (a) blinded written feedback by simulation experts on AI-generated teamwork reports and (b) learner interviews about being observed by AI (Methods). - **n:** 41 participants (16 nurses incl. 11 anesthesia nurses, 17 residents, 8 attendings) (Methods – Participants; Table 1); 26 AI-generated reports from 9 scenarios reviewed by 4 simulation experts; 27 learner interviews (Results). - **Setting:** University Hospital Zurich anesthesia simulation center, Jan 2025; teams of 5–7 clinicians per training day, two airway scenarios, debriefings by two facilitators; lapel-mic transcripts analysed by Isaac (Saipient AI) and ChatGPT-4o (same underlying LLM) with a Team-FIRST prompt; reports produced in <2 min after each scenario (Methods). - **Findings:** 1. Experts valued "broader observation capture": AI transcripts/reports surfaced interaction details experts missed during live observation, and several saw the report as a helpful adjunct for debriefing (Results – experts' theme 1). 2. Limitations: inaccurate/overlapping assignment of quotes to Team-FIRST categories, speaker/role misattribution, misinterpreted technical terms, overconfident wording, possible role/profession bias, no nonverbal or tonal information (Results – experts' themes 2–4). 3. Learners showed optimism tempered by caution (transparency, data protection concerns) and insisted interpretation and debriefing be led by a human (Results – learners' themes; Table 3). - **Relevance:** (a) facilitator use of AI: debriefing facilitators evaluate LLM teamwork reports as preparation aid | (b) AI planning/running/supporting group sessions: LLM analysis of team interaction to support facilitated group debriefings | (c) AI and group ideation/decisions: — - **Caveats:** Single centre; reports assessed by study-team experts (co-authors), not used live in debriefings with learners; qualitative only; no accuracy metrics; clinical simulation context. - **Screening check:** Meets criteria — empirical qualitative study with stated method and n; LLM performs a facilitation-support function on multi-person team interaction data, with facilitators as respondents; professional (CME) training, not individual classroom use; English. #### R1689 — GPT-4o-generated scripts for simulation debriefing (pilot) - **Citation:** Hong, E., Kazmir, S., Dylik, B., Auerbach, M. A., Rosati, M., Athanasopoulou, S., et al. (2025). Exploring the Use of a Large Language Model in Simulation Debriefing: An Observational Simulation-Based Pilot Study. Simulation in Healthcare, 20(6), 366-371. https://doi.org/10.1097/sih.0000000000000861 - **Type:** journal — peer-reviewed: yes - **Full text read:** abstract only (not open access; PubMed page blocked) - **Method:** prospective, observational, simulation-based pilot study; NASA-TLX (primary outcome) and survey-based perception items; all data self-reported (Abstract) - **n:** 4 facilitators and 25 learners (Abstract); number of simulations not verified - **Setting:** Yale University School of Medicine simulations; for each simulation a debriefing script was generated by passing a real-time transcription of the case to GPT-4o, used by facilitators in the debrief (Abstract) - **Findings:** 1. Facilitators and learners showed strong enthusiasm for AI integration, mean Likert 4.75/5 and 4.0/5 respectively (Abstract). 2. NASA-TLX indicated moderate-to-high mental demand for facilitators (reported in the abstract as "M = .8/21; SD = 6.4", apparently a typographical error — exact value not verified) and learners (M = 9.9/21; SD = 4.5) (Abstract). 3. AI perceived to help maintain focus (M = 4.8/5), support learning objectives (M = 4.2/5) and minimize distractions for facilitators (M = 4.6/5) and teams (M = 4.5/5); authors note continued need for human oversight (Abstract). - **Relevance:** (a) facilitators used GPT-4o-generated scripts in live debriefs | (b) LLM support for running team debrief sessions; perceived benefits, task load still considerable | (c) — - **Caveats:** abstract only; very small n (4 facilitators); self-report only; no control condition; healthcare-education setting - **Screening check:** meets criteria — empirical pilot with facilitators using an LLM to support facilitated group debriefing; method and n reported; English. #### R1658 — AI to support simulation debriefing (scoping review) - **Citation:** Alnazawi, A. A., & Almarhabi, M. (2026). Artificial Intelligence to Support Debriefing in Simulation-based Healthcare Education: A Scoping Review. Research Square (preprint). https://doi.org/10.21203/rs.3.rs-9261962/v1 - **Type:** review (scoping review) — peer-reviewed: no (preprint) - **Full text read:** yes (https://www.researchsquare.com/article/rs-9261962/v1.pdf; page numbers as printed "Page X/24") - **Method:** scoping review per Arksey & O'Malley and JBI guidance, reported per PRISMA-ScR; MEDLINE, Scopus, Web of Science, CINAHL, no date restriction, English only (Abstract p. 2; Methods pp. 5-7) - **n:** 46 records identified, 25 duplicates removed, 21 screened, 14 excluded, 7 studies included (2023-2026) (pp. 7-8) - **Setting:** AI supporting (facilitator-led) debriefing in simulation-based healthcare education; team-based and screen-based simulations; nursing students, residents, attendings, interprofessional teams (p. 9) - **Findings:** 1. Three AI domains: communication/performance analytics (speech recognition, NLP); generative AI (LLM) systems producing structured teamwork reports and facilitator feedback/guidance prompts; learner-facing reflective dialogue systems (Abstract p. 2; AI Characteristics p. 9). 2. Reported outcomes were mainly feasibility, usability and perceived educational value, with limited evidence of objective performance improvement; no large-scale experiments or RCTs were identified (Abstract p. 2; p. 9). 3. Across studies AI was positioned as an adjunct supporting rather than replacing the human facilitator (p. 9). - **Relevance:** (a) facilitators use genAI for feedback tools and structured debrief prompts | (b) LLM-generated teamwork reports support facilitator-led team debriefs; evidence mostly feasibility-level | (c) — - **Caveats:** preprint; only 7 studies, largely single-institution; not all included studies use generative AI; overlaps topically with R1689 (likely one of the included studies — not verified) - **Screening check:** meets criteria — scoping review with stated method and n; debriefing is a facilitated group session and generative AI is among the covered systems; English. #### OA2087 — Integrating AI into Citizens' Assemblies (framework + key-informant interviews) - **Citation:** McKinney, S. (2024). Integrating Artificial Intelligence Into Citizens' Assemblies: Benefits, Concerns and Future Pathways. Journal of Deliberative Democracy, 20(1), pp. 1–12. https://doi.org/10.16997/jdd.1556 - **Type:** journal — peer-reviewed: yes - **Full text read:** yes (https://delibdemjournal.org/article/1556/galley/5278/download/, 12-page PDF) - **Method:** Mainly conceptual: an analytical framework applied to 11 possible AI/LLM applications across a citizens' assembly (CA) lifecycle, informed by 4 key-informant interviews, the Global Assembly on the Climate and Ecological Crisis as a case study, and literature; "applied normative theorising" (p. 1–2, Methodology). - **n:** 4 key-informant interviews with CA practitioners (Kelly McBride, Yago Bermejo, Claire Mellier, Johnny Stormonth-Darling) (p. 2). No AI was deployed; no participant data. - **Setting:** Practitioners who design and deliver national, transnational and global CAs; discussion of hypothetical LLM uses (e.g., GPT-4-class models) for translation, summarisation, Q&A, facilitation, aggregation and consensus statements. - **Findings:** 1. Interviewees responded positively to using LLMs to simplify and summarise information for participants. Two interviewees (Stormonth-Darling, Bermejo) repeatedly stressed human-in-the-loop oversight to handle bias and hallucination, and said LLMs could not be trusted in knowledge-provision roles without fine-tuning and human verification (p. 4, p. 6). 2. On AI facilitation, "as McBride and Bermejo stressed, AI facilitation will not be as effective as human facilitation", because the facilitator's sensitivity to inequalities and nuance is "an importantly human craft". The author suggests hybrid models, e.g., rotating groups between AI and human facilitators (p. 7). 3. On LLM-generated consensus statements, McBride asked "at what point does the AI become a participant?" The author concludes that machine-generated ideas must be critically deliberated rather than accepted uncritically, and that AI aggregation is less transparent than human whiteboard work (p. 8). - **Relevance:** (a) facilitator use of AI: practitioner views on where LLMs could help or harm CA facilitation | (b) AI planning/running/supporting group sessions: hypothetical mapping only, with no deployment | (c) AI and group ideation/decisions: concerns about undue AI influence on consensus - **Caveats:** Mostly conceptual; n=4 named experts; interview findings reported selectively as quotes; no systematic qualitative analysis described. - **Screening check:** Borderline PASS: meets the criterion "facilitators/meeting professionals as respondents" (4 CA practitioners) and reports method and n, but the empirical part is thin and mostly about hypothetical AI use. The screener may reclassify it as FAILS (primarily conceptual). #### OA2462 — Urban planners and a WhatsApp LLM chatbot for AI-mediated participation (Bogotá region) - **Citation:** Dortheimer, J., Oren, M. L., & Sprecher, A. (2026). Automating Consultation? Urban Planners and Chatbot-Based AI-Mediated Participation. Journal of Urban Technology, 1–25. https://doi.org/10.1080/10630732.2026.2693907 - **Type:** journal — peer-reviewed: yes - **Full text read:** abstract only (closed access per OpenAlex; Taylor & Francis returned HTTP 403; abstract from OpenAlex) - **Method:** Two-year embedded case study; semi-structured interviews with planners (abstract). Number of interviews not verified. - **n:** 273 citizen-to-AI interactions across 2 deployments (abstract). Number of planners interviewed: not verified. - **Setting:** Regional strategic planning process in the Sabana Occidente region of Bogotá, Colombia. A WhatsApp-based LLM chatbot handled visioning, project exploration and voting (abstract). Model not verified. - **Findings:** 1. Planners reported that the chatbot "reduced participation barriers but raised concerns about misinformation and shallow depth of deliberation" (abstract). 2. It "expanded access and supported prioritization, yet remained closer to consultation than deliberation" (abstract). 3. Democratic value "depended primarily on governance arrangements—validation, accountable summarization, and traceability—rather than interface novelty" (abstract). - **Relevance:** (a) facilitator use of AI: planners (engagement professionals) on using an LLM chatbot for consultation | (b) AI planning/running/supporting group sessions: the chatbot replaced facilitated engagement with 1:1 citizen–AI chats | (c) AI and group ideation/decisions: supported prioritisation, but deliberation was shallow - **Caveats:** Abstract only; citizen–AI interactions were individual (1:1), not group sessions; the planner interview n is unknown. - **Screening check:** Borderline. It meets inclusion only through the "facilitators/meeting professionals as respondents" route (planners interviewed about AI-mediated participation). The AI interactions themselves are not group sessions, and n (interviews) cannot be confirmed from the abstract. Flag for verification; FAILS if the planner interviews are not reported with an n. #### OA2019 — Questionable devices: ChatGPT in carbon-removal deliberative workshops - **Citation:** Waller, L., Moats, D., Cox, E., & Bellamy, R. (2024). Questionable devices: Applying a large language model to deliberate carbon removal. Environmental Science & Policy, 162, 103940. https://doi.org/10.1016/j.envsci.2024.103940 - **Type:** journal — peer-reviewed: yes - **Full text read:** yes (https://kclpure.kcl.ac.uk/ws/files/321979788/1-s2.0-S1462901124002740-main.pdf, publisher PDF, 9 pp.) - **Method:** Qualitative study of deliberative workshops; open coding (NVIVO) of workshop transcripts, "loosely inspired by situational analysis"; authors "make no strong claims about our inductive findings" (p. 4, fn. 11). - **n:** 4 day-long workshops (London, Cardiff, Edinburgh, Belfast), each split into 2 groups = 8 deliberative groups of "~8 participants" each (p. 3, incl. fn. 9). Total participant count not stated explicitly (8 x ~8 would be ~64; not verified). - **Setting:** UK public recruited topic-blind via an agency for sociodemographic diversity; moderated deliberation on land-based carbon removal. ChatGPT (version 3.5) used in two sessions: groups formulated and agreed one question to pose to ChatGPT at the start, then read and discussed a printout of the response at the end; moderators stayed "agnostic" and turned questions back to the group (p. 3–4). - **Findings:** 1. Groups mostly reached their question by straw-poll voting under time pressure; questioning was "predominantly expressed as a technical activity of 'querying'"; almost all participants knew ChatGPT and 7 of 8 groups contained at least one active user, who coached others on prompt wording (p. 4, §3 and §4.1.1). 2. All but one of the 8 ChatGPT texts came back as lists; in almost all groups the lists were appraised as a "summary"/"overview"/"comprehensive", and in all but one workshop the common view was that the lists were a "starting point" for further inquiry; one group found its text "vague"/"lacking detail", and one participant dismissed it as "just words on paper" (p. 5, §4.1.2, fn. 14). 3. Overall, ChatGPT "was all too often engaged unquestioningly as an instrument for informing discussion"; direct challenges (e.g., on commercial bias) were rare, and in one of them the moderator was pulled into speaking for the tool's designers (Abstract p. 1; p. 4–5, §4.1.1). - **Relevance:** (a) facilitator use of AI: moderators introduced ChatGPT as a scripted deliberation device and had trouble staying neutral when participants challenged it | (b) AI planning/running/supporting group sessions: the LLM served as an information input to moderated workshops, with qualitative evidence of how groups framed and appraised its output | (c) AI and group ideation/decisions: groups tended to treat LLM lists as authoritative summaries, which narrowed their questioning - **Caveats:** Interpretive STS analysis with no outcome metrics; 2023-era GPT-3.5; total n not reported explicitly. - **Screening check:** PASSES (empirical; 8 human groups deliberating with LLM involvement; method and group sizes reported; English). #### OA2366 — AI-supported youth citizen science (facilitator-mediated GPT + image AI) - **Citation:** De la Vega-Taboada, E., Portillo, S. A., Gómez-García, L. M., Banchoff, A., Bermudez, V. M., Chavez, D. M., Sgaraglino, T. I., Flores Millender, E., Sarmiento, O. L., & King, A. C. (2026). A New Model for Youth-Driven Community Change: Exploratory Testing of Artificial Intelligence–Supported Citizen Science. JMIR AI, 5, e79464. https://doi.org/10.2196/79464 - **Type:** journal — peer-reviewed: yes - **Full text read:** yes (Europe PMC full-text XML, PMC13173061; the article has no page numbers, so sections are cited) - **Method:** Exploratory action research in two phases: (1) researcher-only prompt testing (3 rounds of SecureGPT prompts across stakeholder roles; DALL-E vs Adobe Photoshop AI); (2) implementation in the Our Voice process. Descriptive, process-oriented analysis of field notes, analytic memos, archived prompts and outputs, and session audio recordings (Abstract: Methods; Methods: Data Collection and Analysis During Implementation). - **n:** 12 adolescent citizen scientists (7 girls, 5 boys; mean age 14, SD 1.41) from a public school in El Pozón, Cartagena; 38 decision makers attended the advocacy meeting (Methods: Participants and Recruitment; Results: The Our Voice Advocacy Step). - **Setting:** Facilitated Our Voice "Discuss" and "Advocacy" meetings, conducted in Spanish. AI use was facilitator-mediated: youth co-developed prompts by facilitated consensus, a researcher-facilitator entered them into SecureGPT (a privacy-enhanced ChatGPT 4.0) and Adobe Photoshop AI, and outputs were read aloud and reviewed with the group in real time (Methods: Data Collection and Analysis During Implementation). - **Findings:** 1. Prompt testing: role-based prompts gave redundant responses for similar roles. Question-oriented prompts drew "more thoughtful and reflective responses" than direct feedback. A fixed "3 strengths, 3 weaknesses, 5 questions" format was chosen after overly long outputs (Results: Preparation Phase). 2. In sessions, GPT feedback led the group to revise solutions. For waste, they moved from awareness-raising alone to containers plus education; for stagnant water, from filling holes with sand/cement to removable cement lids. The facilitator "clarified key points and corrected technical errors when they arose" (Results: Waste Management; Stagnant Water). 3. DALL-E did not produce realistic place-based edits, while Photoshop AI with iterative prompting did. The AI-refined proposals and images, shown to 38 decision makers, led to committees and 1- and 3-month commitments (e.g., containers in the school, lids for planter holes). Highly realistic images created a "feasibility gap" and had to be framed as aspirational (Abstract: Results; Results: The Our Voice Advocacy Step). - **Relevance:** (a) facilitator use of AI: a facilitator operated GPT and image AI live as a structured deliberation aid with a concrete workflow | (b) AI planning/running/supporting group sessions: GPT critique fed into facilitated group problem-solving meetings | (c) AI and group ideation/decisions: AI feedback prompted the group to refine solutions toward more feasible ones (qualitative) - **Caveats:** n=12, single site, no comparison condition; researcher-authored descriptive analysis; the stakeholder commitments cannot be attributed to AI. - **Screening check:** PASSES (empirical; a youth group deliberating with facilitator-mediated LLM use; method and n reported). #### S2-345 — Fast Track for Collaborative Solutions (LLM-supported public service design workshop) - **Citation:** da Silva, W. C. M. P., Canedo, E. D., Marsicano Correa, G., de Moura, H. G., Franco, P. E. D., Santana, G. G. N., & Figueiredo, R. M. da C. (2025). Fast Track for Collaborative Solutions: A Methodology for Public Service Design Workshops Supported by LLMs. In dg.o 2025 – 26th Annual International Conference on Digital Government Research, 9–12 June 2025, Porto Alegre, Brazil (15 pp.). https://doi.org/10.59490/dgo.2025.1061 - **Type:** conference (research paper) — peer-reviewed: yes (submitted/revised/accepted dates given, p. 1) - **Full text read:** yes (publisher PDF, CC-BY-4.0, pdf3/S2-345.pdf; pages cited as "Page X of 15") - **Method:** Single field case study / methodology proposal: one two-day service design workshop in which LLMs were used for idea consolidation, consensus facilitation and documentation; evaluation "based on participant feedback, the quality of the MVP, and the efficiency of the workshop process", with stated metrics "time required for each stage, the perceived effectiveness of LLM integration, and the alignment of the final prototype" (p. 5). No results for these metrics are reported. - **n:** 28 participants (public servants, managers, technical staff) from various roles and public organisations; 1 workshop over two days; small groups (number not stated); 48 post-it insights generated (pp. 4, 6). - **Setting:** Brazil, "Startup EstataisGov" workshop with SEST (Secretariat of Coordination and Governance of State-Owned Enterprises) on redesigning the Functions Plan submission/analysis service; stages: problem identification, Is/Is-Not solution exploration, personas, user journeys, low-fi prototypes, MVP (pp. 4–5, 10). LLM(s) unnamed; participants submitted contributions via "a platform integrated with the LLM"; example prompts reported (p. 5, p. 6). - **Findings:** 1. Time claim, exactly as stated: "Despite identifying some challenges during the public service design workshops supported by LLMs, several key advantages were observed: 1) Time Reduction: The implementation of the Fast Track reduced the duration of workshops by at least 30%" (§5 Discussions, p. 12). Basis: none stated. The paper reports no baseline/comparison workshop, no timing data per stage, and no description of how the 30% was measured; the abstract states only that the methodology "significantly reduced execution time ... without compromising decision quality" (p. 1). 2. LLM categorised the 48 post-it insights and, via a follow-up prompt, consolidated them into seven problem categories and a core problem statement, validated by participants; accuracy described qualitatively only: "Overall, the accuracy level was high, as the categories identified by the LLM aligned closely with the responses provided by stakeholders" (p. 6). 3. An LLM-generated persona was not shown to participants but compared with participant-created personas; differences were found "particularly concerning the 'pain points' highlighted by more experienced stakeholders" (p. 8). Other claimed advantages (improved collaboration, "more informed decisions") are asserted without data (p. 12). - **Relevance:** (a) facilitator use of AI: facilitators/researchers used LLM prompts to cluster, consolidate, generate voting options and summarise in a live workshop | (b) AI planning/running/supporting group sessions: LLM support across a two-day design workshop (consolidation, consensus options, documentation) | (c) AI and group ideation/decisions: claims of faster consensus and unchanged decision quality, but unmeasured - **Caveats:** Single case, no control or baseline; LLM model not named; efficiency/decision-quality claims unsupported by reported data; the Threats to Validity section refers to "surveys, interviews, and system logs" and results being "statistically significant" (p. 12), none of which are reported in the paper; authors are the workshop organisers. - **Screening check:** Meets criteria (weakly): empirical field case, 28 humans in a group workshop with LLM involvement, method and n reported, English. Evidence quality for outcomes is very low. #### S2-416 — Midjourney in community consultation workshops, Samarinda - **Citation:** Nikolaeva, K., Guaralda, M., & Nyandega, D. (2026). Imagining climate-resilient urban spaces: Integrating AI into community consultation in Samarinda. Interaction Design and Architecture(s) Journal (IxD&A), N.68, 157–191. https://doi.org/10.55612/s-5002-068-006 - **Type:** journal — peer-reviewed: yes (the text does not state its review process; IxD&A is a peer-reviewed journal) - **Full text read:** yes (local PDF pdf3/S2-416.pdf, via https://doi.org/10.55612/s-5002-068-006) - **Method:** Action research / single case study. The authors were both researchers and workshop facilitators. Design–review cycles ran Jul–Nov 2023: an internal tool comparison (DALL·E, Stable Diffusion, Midjourney), an online expert session, three in-person workshop days in Samarinda (24–27 Sep 2023) and a half-day expert synthesis workshop at QUT Brisbane (28 Nov 2023). Artefacts (sketches, prompt fragments, 133 images) were analysed with inductive reflexive thematic analysis plus KJ affinity diagramming (pp. 163–164, 173, 176; §3, §3.4). There are no questionnaires and no quantitative outcome measures. - **n:** Samarinda: "Each day engaged approximately 40- 45 participants" (p. 165–166, §3.3); "In total, roughly 130 participants took part across the three days" (p. 166, §3.3). This is the exact wording; there is no exact headcount. Participants worked in "small focus groups of about five… each guided by one facilitator" (p. 166). The next page says "Three facilitators and local partners operated per day (one per group)" (p. 167), which does not square with ~40–45 people in groups of ~5, and the text does not resolve this. The number of groups is not stated (Fig. 3 mentions a "Group 9"). Artefacts: 133 Midjourney images and "more than one hundred prompt components" (p. 180, §4). Brisbane expert workshop: six participants in two groups (pp. 176–177, §3.4). - **Setting:** Indonesia (Samarinda, East Kalimantan), within an Adaptation Fund climate-adaptation project. The groups co-designed a flood-adaptive multi-purpose public space. Day 1 was local community members, Day 2 civil-society organisations and market traders, Day 3 city-government officials, planners and academics (p. 166). The workshops ran in Bahasa Indonesia, and prompts were translated into English for the model. Each facilitator's PC ran Midjourney v5.2 through Discord on a large shared screen, one per ~5 participants. The day's sequence was: analogue exercises (dream map, symbols & values, verbal brainstorm), then a 45-min group prompt-crafting session with a 5-question template, then image generation and refinement, then a plenary review (pp. 167–172, §3.3). - **Findings:** 1. Participants learned to prompt over the three days. On Day 1, newcomers wrote "broad, relatively simple prompts and showed some hesitation". By Day 3, professionals wrote "more specific, multi component prompts". "By the end of the workshops, participants from all backgrounds demonstrated basic literacy in how AI image generation operates" (pp. 180–181, §4). All of this is qualitative facilitator observation, not measured. 2. The prompt–image feedback loop "surfaced unspoken priorities". Groups added cultural markers, a "small mosque corner" and a timber material palette after seeing images that were "too foreign", and asked for privacy buffers between toilets and prayer spaces (p. 181, §4). Coding of the prompts and images gave nine themes, clustered into three: Human-Centric Design, Environmental Integration, Socio-Cultural Cohesion (pp. 173, 176, 180). 3. The authors attribute the co-creation to the facilitated protocol, not to the AI. They write that generative models "do not automatically create participation, influence, or accountability", while Midjourney "compresses the time and skill required" to visualise ideas (p. 182, §5). Reported challenges: prompt literacy needed a cheat sheet and real-time coaching; opaque moderation blocked "a locally significant term"; photorealism led some participants to mistake images for existing projects, so facilitators switched to sketch or watercolour styles; there was non-local bias in outputs; and access costs (Discord, paid plan) (pp. 183–184, §5). - **Relevance:** (a) Authors acting as facilitators embedded Midjourney in a workshop protocol, coached prompting, curated images and managed moderation and realism bias | (b) Generative AI supported live small-group co-design workshops as a visualisation and boundary-object tool | (c) Qualitative claim that AI imagery broadened the range of ideas and surfaced tacit priorities; no controlled comparison. - **Caveats:** Participant counts are approximate ("roughly 130"; ~40–45 per day). The group and facilitator counts are internally inconsistent (p. 166 vs p. 167). There was no control condition and no participant survey. Researchers were also the facilitators and the analysts. Claims such as "Many reported thinking more like designers" are not quantified. The image model is v5.2 (2023). - **Screening check:** Meets the criteria. It is empirical, with small groups (~5 people) interacting with generative AI (Midjourney) in facilitated workshops, and it reports method and (approximate) n. English. Note that the AI is a text-to-image model, not an LLM; the criterion says "generative AI/LLM", so it qualifies. #### S2-538 — WeDesign: Generative AI–Facilitated Community Consultations - **Citation:** Mushkani, R., Berard, H., & Koseki, S. (2025). WeDesign: Generative AI–Facilitated Community Consultations for Urban Public Space Design. arXiv preprint arXiv:2508.19256 (v3, 15 Sep 2025). https://doi.org/10.48550/arXiv.2508.19256 - **Type:** preprint — peer-reviewed: no (preprint) - **Full text read:** yes (https://arxiv.org/abs/2508.19256, v3; local pdf3/S2-538.txt) - **Method:** exploratory "mixed qualitative approach": one in-person half-day focus-group workshop + six semi-structured expert interviews; open coding (two coders on 20% of data, Cohen's κ computed but values not reported), LDA topic modelling on transcripts/notes/prompts, word cloud; no inferential statistics ("given the exploratory design", p. 7 fn. 1) (§3, pp. 4–7). - **n:** 29 workshop attendees = 14 citizens + 5 computer scientists + 5 urban architects + 5 note takers (§3, p. 4); participants grouped into 5 teams, each with 1 AI specialist, 1 urban professional and 2–3 citizens (§3.3, p. 6) — i.e., the 29 includes the 5 note takers, and the stated team composition accounts for 24 people (5 + 5 + 14). The **14-person identity diagram** (Figure 1, pp. 4–5) "depicts the self-reported identity markers of the 14 participants", i.e., the 14 citizen participants recruited via outreach ("a workshop cohort of 14 participants who self-identified across various demographic markers", §3.1, p. 5); it does not cover the AI specialists, urban professionals or note takers. Plus 6 interviews (~1 h each) with professionals in municipal facilitation, architectural/urban design and parapublic planning (§3.2, pp. 5–6). - **Setting:** Montréal, bilingual (French/English); WeDesign platform built on Stable Diffusion XL (text-to-image, not an LLM); five public-space scenarios (A–E); ~30 min brainstorming then ~3 h of prompting and iterative refinement; ~440 prompts, 4 images each; bilingual facilitation; AI specialist in each group guided prompt syntax (§3.3, p. 6; §3.4, p. 8). - **Findings:** 1. Real-time images increased engagement: participants described the workshop as more engaging and less "dry" than conventional presentations ("it feels like we're all brainstorming together ... instead of just listening to a lecture"); serendipitous image elements (e.g., an unprompted wooden bench) seeded new ideas (§4.6, p. 10). Qualitative only; the descriptive engagement metrics announced in §3.4 (prompts per role, hearts, speaking turns) are not reported in the results. 2. Representational limits: repeated prompting for diversity/accessibility yielded tokenistic elements (e.g., "random wheelchairs everywhere") and rarely functional accessibility features; French-speaking participants prompted in English because SDXL performed better in English (§4.2, p. 8; §4.4–4.5, p. 9). 3. Facilitation-relevant requirements: participants and experts requested region-specific in-painting, multilingual prompt support, built-in group voting "rather than relying on ad hoc verbal consensus", and preference sliders; experts warned that "an infinity of images can create expectations too high" and that the approach needs structured facilitation and expectation management (§4.7–4.8, pp. 11–13; §5, p. 14). - **Relevance:** (a) interviews include municipal facilitation professionals' views on generative AI in consultations | (b) generative AI (text-to-image) as a shared tool within facilitated in-person focus groups | (c) images stimulated group ideation but polished outputs risk overshadowing feasibility; no decision-quality measure - **Caveats:** Preprint; single workshop, single city, self-selected residents; text-to-image model (SDXL), not an LLM; purely qualitative/descriptive (κ values and engagement metrics not reported); authors are also the platform developers; sample count inconsistency (29 attendees including 5 note takers vs. 14-person identity figure) should be reported as above. - **Screening check:** Meets criteria — empirical; groups of ≥2 humans interacting with generative AI in a facilitated consultation; also professionals (incl. municipal facilitation) as interview respondents; method and n reported; English. Note: generative AI is text-to-image, not LLM — acceptable only if the review's "generative AI" scope includes image models. #### S2-473 — Designing expressive-arts family workshops using AI tools - **Citation:** Tiwari, S. (2025). Designing Expressive Arts Workshops for Families Using AI Tools. International Journal of Designs for Learning, 16(2), 194–204. https://doi.org/10.14434/ijdl.v16i2.42188 - **Type:** journal (design case) — peer-reviewed: unclear (IJDL is a peer-reviewed AECT journal, but the paper itself does not state its review process) - **Full text read:** yes (local PDF pdf3/S2-473.pdf, https://doi.org/10.14434/ijdl.v16i2.42188) - **Method:** Single-author, first-person design case of two iterations of an online family expressive-arts workshop ("Hi, Oki Pie!", Summer 2025), designed and facilitated by the author. Evidence comes from post-workshop family surveys (instruments in the Appendix, p. 204) and the facilitator's reflections and observations. Design feedback came from a registered art therapist, makerspace board members and a Stanford d.school "Edtech Remix" cohort (p. 195). There is no formal analysis method and no statistics. - **n:** 7 families (Table 1, p. 195): caregivers were parents or grandparents aged 28–70, children aged 4–8 per Table 1. The text also says "in the age range of 4-10" (p. 195), which is inconsistent with the table. 4 of the 7 families had attended earlier workshops. Two workshop sessions (iterations). The surveys report "out of 7 families" for both iterations (pp. 199, 201), which implies that the same 7 families attended both sessions; this is not stated explicitly. The number of individual participants is not stated. - **Setting:** Online group workshop (shared screen, chat, Padlet) for families in the San Francisco Bay Area. One facilitator, the author (character designer working towards expressive-arts REACE registration). Iteration 1 used four AI tools: a Playlab-built "Playtime Brainstormer", Brickit, a custom ChatGPT GPT ("Oki Pie's Toy-to-Story Generator", which the facilitator ran on shared screen with families' LEGO photos), and Meta's Animated Drawings. Iteration 2 used one AI tool, Drawalive, with the facilitator generating the videos (pp. 196–202; Table 2, p. 202). - **Findings:** 1. Iteration 1 (four AI tools): "5 out of 7 families responded that they enjoyed the workshop 'a lot' overall… However, 6 out of 7 families rated the mix of AI tools and hands-on art as 'too much tech' and noted that the activities were only 'somewhat' accessible" (p. 199). The author names three design failures: too many tools, too few breaks, and "challenging facilitation due to too many task-switches" (p. 199). 2. Iteration 2 (one AI tool): "all families liked this change in workshop structure to include only one AI tool. 4 out of 7 families preferred just the free play activities, whereas 3 families preferred a combination of both" (p. 201). Families described the balance as "just right" (Table 2, p. 202). 3. Facilitator role: in iteration 1 "my focus was on timekeeping and troubleshooting". In iteration 2 "I felt more relaxed and present… focused on facilitation and observations" (Table 2, p. 202). The author's role evolved "from being a timekeeper and troubleshooter to being a facilitator and observer" (p. 203). - **Relevance:** (a) A facilitator designs custom GPT and Playlab tools and integrates AI tools into group workshop design; reflects on the facilitation load that multiple AI tools impose | (b) AI tools used within live online group (family) workshops; more tools meant higher overload | (c) — (no measure of ideation or decision quality; creative outputs are described only). - **Caveats:** Very small n (7 families). Self-report surveys with descriptive counts only. The author is the designer, facilitator and evaluator at once. The "group" consists of families (children aged 4–8 plus caregivers) doing expressive-arts play, not a professional workshop or meeting. Borderline fit. - **Screening check:** Borderline, but meets the criteria. It is empirical (surveys plus observation), with a facilitated online group session in which multiple families interact with generative AI tools (a custom ChatGPT GPT, a Playlab LLM app, Drawalive), and it reports method and n (7 families, two sessions). English. The facilitator perspective is directly reported. #### S2-324 — AI-generated vignettes as triggers for collaborative reflection - **Citation:** Kaup, C. F., Larsen, J. D., Kjær, R. S., & Dau, S. (2025). Artificial Intelligence-Generated Vignettes as Triggers for Collaborative Reflection: Exploring Methodological Potentials in Higher Education. Academic Quarter | Akademisk kvarter, 31 (Autumn 2025), 92–110. https://doi.org/10.54337/academicquarter.i31.11273 - **Type:** journal — peer-reviewed: unclear (not stated in the PDF) - **Full text read:** yes (publisher PDF, pdf3/S2-324.pdf; printed page numbers = PDF page + 91) - **Method:** Qualitative case study (Yin); data = written group reflections + one audio-recorded focus-group interview, both structured by the same question sheet; deductive analysis using Akkerman & Bakker's four boundary-learning mechanisms and Schön's reflection-in-action; pattern matching; selected transcripts independently reviewed by a second researcher; single data-collection point (pp. 96–99). - **n:** 23 pre-service social educators (3rd semester, school and leisure specialisation, Danish university college) invited; 3 in a focus group interview, the rest in 6 reflection groups of 3–4 students (pp. 97–98; Table 1 p. 98). - **Setting:** Course module; the research team produced a short GenAI video vignette (script via "a generative language model", video via "a text-to-video AI technology"/"multimodal GenAI platform"; specific tools not named) through iterative prompting, depicting an AI emotion-recognition app for autistic children; groups discussed it using a 5-question sheet (pp. 96–97; Appendix pp. 108–110). - **Findings:** 1. The vignette functioned as a "boundary object": groups moved from initial normative rejection (identification) to negotiating divergent interpretations (coordination) and reconsidering assumptions (reflection) (pp. 100–103). Example: one group first wrote "It's a bad idea [...] because it takes away the personal aspect of relationships" and later proposed "You could use it as a training tool" (Reflection, Group 1) (p. 103). 2. Disagreement was sustained productively rather than pushed to consensus: "It was more legitimate to disagree [...] and find solutions" (Group 2); "We had a good discussion [...] and all came up with relevant ideas" (Group 1) (p. 103). 3. Design recommendation: GenAI vignettes should be designed for ethical ambiguity, recognisability and affective resonance, "and facilitated in ways that keep multiple interpretations in play rather than push for early consensus" (p. 104). - **Relevance:** (a) facilitator use of AI: educators use GenAI to prepare session stimulus material (video vignette) for group reflection | (b) AI planning/running/supporting group sessions: GenAI-made trigger material for a facilitated group discussion; AI does not act during the session | (c) AI and group ideation/decisions: qualitative account of group reflection/perspective shifts; no quantitative outcomes - **Caveats:** Very small qualitative sample, single session, no comparison (e.g., text vignette) condition; GenAI tools not named; AI involvement limited to pre-produced material; claims of transformation rest on selected quotes; who facilitated the groups/focus group is not described. - **Screening check:** Borderline pass: empirical, groups of 3–4 humans discussing GenAI-produced session material (GenAI involvement via facilitator-prepared stimulus), method and n reported, English. If the review requires AI to be active during the group interaction, this would FAIL on that criterion. #### S2-303 — GenAI to counter learner groupthink via controversy (CALIE) - **Citation:** Wiss, A., Showstark, M., Dobbeck, K., Pattershall-Geide, J., Zschaebitz, E., Joosten-Hagye, D., Potter, K., & Embry, E. (2025). Utilizing Generative AI to Counter Learner Groupthink by Introducing Controversy in Collaborative Problem Based Learning Settings. Online Learning, 29(3), 39–65. https://doi.org/10.24059/olj.v29i3.5096 - **Type:** journal — peer-reviewed: yes - **Full text read:** yes (publisher PDF, pdf3/S2-303.pdf; printed page numbers = PDF page + 38) - **Method:** Mixed-methods, single-arm (no control group) post-activity survey: 3 Likert items + yes/no facilitator-influence item analysed with repeated-measures ANOVA, MANOVA and correlations; open-text comments coded with categorical/open coding (pp. 46–47). IRB: Yale #2000024050 (p. 43). - **n:** 165 graduate health-professions learners took part; 158 opted into the study/survey (p. 43). Twelve small groups, each led by one faculty facilitator (p. 43); elsewhere described as "small groups of 15 students" (p. 44). 7 graduate programs (p. 43). ANOVA df (1, 156) (p. 47). - **Setting:** US virtual interprofessional education (VIPE) session: 180-min synchronous problem-based learning case on the opioid crisis, with a 60-min faculty-facilitated small-group segment (pp. 44–45). "CALIE" was built by faculty in ChatGPT 4.0, prompted to act as a health services expert; outputs were pre-generated and selected by faculty (prompted for "potentially controversial responses"), inserted into the Facilitator's Guide and read aloud by the facilitator at three points ("I will 'voice' for a generative artificial intelligence named CALIE") (pp. 44–45). Students were not told CALIE was intentionally placed as a groupthink disruptor (p. 43). - **Findings:** 1. Perceptions of AI differed across items, F(2, 156) = 26.01, p < .001: "AI can be a useful tool" M = 3.49 > "AI's feedback was helpful" M = 3.21 > "perceived AI as part of the team" M = 2.88; all pairwise ps < .001 (p. 47; Table 3 p. 48). 2. Facilitator influence: students who reported that their facilitator influenced how the group responded to and integrated the AI rated it higher on all three items: part of team 3.08 vs 2.64, F(1,156) = 5.96, p = .02; useful tool 3.67 vs 3.26, F(1,156) = 5.70, p = .02; feedback helpful 3.44 vs 2.93, F(1,156) = 10.26, p = .002 (p. 47; Table 3 p. 48). The authors did not capture whether that influence was positive or negative (p. 56). 3. Qualitative: the AI's contrarian outputs "helped to stimulate diverse conversations" and prompted group ideation; groups often "coalesced its thinking in opposition to the AI" ("we disagreed as a group with most of the AI responses and that consensus helped solidify our responses") (pp. 54–55; Table 5 pp. 49–54). Students also saw the AI as disconnected from the discussion ("came up with their answers in a vacuum"), lacking empathy/lived experience, and biased (pp. 54–55). Authors state groupthink indicators "were not observed" (pp. 56–57). - **Relevance:** (a) facilitator use of AI: faculty pre-generate GenAI "devil's advocate" prompts and voice them in facilitated groups; facilitator influence shapes acceptance | (b) AI planning/running/supporting group sessions: GenAI content injected as a structured facilitation intervention | (c) AI and group ideation/decisions: self-reported stimulation of divergent discussion; no measured decision-quality outcome - **Caveats:** No control group; groupthink was not measured directly (conclusion drawn from survey comments); AI was not interactive (pre-generated, read aloud; authors list this as main limitation, p. 57); 60-min exposure, novelty effect (p. 57). The appendix survey codes 1 = Strongly agree ... 5 = Strongly disagree (pp. 64–65), yet the text reads higher means as "greater agreement"; whether items were reverse-coded is not stated, so the direction of the means is not verified. Group size inconsistency (12 groups vs "groups of 15" for 165 students). Printed DOI on p. 40 has a typo ("doi.org.10..."). - **Screening check:** Meets criteria: empirical, faculty-facilitated human groups with GenAI (ChatGPT 4.0) involvement, method and n reported, English. Borderline on "interacting with" the AI (AI output was mediated by the facilitator), but the AI output is part of the group process. #### C288 — Collaborative GenAI agents in synchronous group settings (speculative team workshops) - **Citation:** Johnson, J. G., Peralta, M., Kaur, M., Huang, R., Zhao, S. G., Guan, R., Rajaram, S., & Nebeling, M. (2025). Exploring Collaborative GenAI Agents in Synchronous Group Settings: Eliciting Team Perceptions and Design Considerations for the Future of Work. Proceedings of the ACM on Human-Computer Interaction, 9(7), 1–33. https://doi.org/10.1145/3757595 (preprint arXiv:2504.14779) - **Type:** journal (PACM HCI / CSCW) — peer-reviewed: yes - **Full text read:** yes (arXiv v1, https://arxiv.org/pdf/2504.14779; page numbers below refer to the arXiv PDF, not the PACM pagination) - **Method:** Exploratory qualitative design study: 2-hour speculative design workshops per team (priming, 5-min group discussion with a mixed-reality "provotype" in collective mode + 5 min in individual side-channel mode, 40-min "Lego Assembly Line" participatory design, reflection worksheets), plus individual semi-structured follow-up interviews (30–45 min) 3–7 days later; thematic coding in Dovetail by 2 then 4 researchers (§3, pp. 6–11). - **n:** 25 professionals in 6 real teams (3–6 members each), 6 workshops; 19 of 25 completed follow-up interviews (§3.2, p. 9; Table 1, p. 8). - **Setting:** Local North American professional teams (XR innovation team, non-profit board, university staff teams, student organisation, AI-policy/legal working group) meeting on a neutral discussion topic (smart street lights with cameras in their city). MR headsets; embodied humanoid GenAI agents (technology, sustainability, ethics expert personas), one per participant; speech content from a "GPT-based pipeline", but a study facilitator used a Wizard-of-Oz interface to control when agents spoke or raised their hand in collective mode (§3.1, pp. 6–8). - **Findings:** 1. All participants said collaborative agents should augment, not replace, existing team skills; teams saw three opportunities for synchronous group sessions such as brainstorming or deliberation: including diverse perspectives / devil's advocate to boost confidence in outcomes, challenging groupthink and drawing on organisational memory, and bridging communication gaps / reducing social friction (e.g., an agent acting "as a neutral moderator") (§4, pp. 12–13). 2. Willingness to integrate agents depended on individual factors (trust, perceived usefulness), team factors (social sensitivity, e.g. knowing when to interrupt; fit to team culture) and organisational/societal factors (§5.1–5.3, pp. 13–16). 3. Three design tensions: humanoid/expressive agents seem more intelligent but may foster misplaced trust; higher social prominence raises engagement but may encourage overuse and undue influence on team outcomes; shared agents make errors easier to detect, whereas private/sub-group agents let humans filter AI contributions and manage its influence (§6, pp. 17–20; §7, p. 21). - **Relevance:** (a) facilitator use of AI: — (no facilitator respondents; a researcher acted as WoZ "facilitator") | (b) AI planning/running/supporting group sessions: professionals' expectations for AI agents participating in meetings, incl. neutral-moderator role and when to interject | (c) AI and group ideation/decisions: perceived value for challenging groupthink and broadening deliberation, and risk of undue AI influence on team outcomes (perceptions only, not measured) - **Caveats:** Speculative/perception study, no outcome measurement; agent turn-taking was Wizard-of-Oz; only 5+5 minutes of actual agent interaction per team; convenience recruitment via university and authors' networks; published PACM version not read (arXiv v1 used). - **Screening check:** Meets criteria — empirical (workshops + interviews), groups of 3–6 humans interacting with LLM-generated agent contributions, method and n reported, English. #### OA959 — GenAI across divergence/convergence in team design thinking (innovation sprints) - **Citation:** Hutter, K., Rupp, J., & Füller, J. (2026). From divergence to convergence: GenAI's role in collaborative design thinking. Business Horizons (in press/online 2026; volume and pages not verified). https://doi.org/10.1016/j.bushor.2026.04.001 - **Type:** journal — peer-reviewed: yes - **Full text read:** abstract only (listed as hybrid OA but ScienceDirect returned 403; abstract via OpenAlex) - **Method:** Qualitative study of three innovation sprints; analysis of GenAI use across problem understanding, ideation and prototyping; derivation of a 2×2 typology (abstract). Specific data sources (interviews, observation, artefacts) not verified. - **n:** 71 interdisciplinary teams and 18 corporate partners across three innovation sprints (abstract); number of individuals not verified. - **Setting:** Innovation sprints with corporate partners (likely university-industry; not verified); GenAI tools not verified. - **Findings:** 1. GenAI "effectively expands the problem space and generates diverse solutions", but integration "varies significantly across teams" (abstract). 2. Critical tension: GenAI can enhance early-stage divergence, but its output "often lacks the contextual and empathic depth required for user-centered innovation" (abstract). 3. 2×2 framework: Level of Integration (High vs Low) × Purpose of Use (Supportive vs Directive), yielding four engagement patterns; practical guidance on "team orchestration and intervention strategies" (abstract). - **Relevance:** (a) facilitator use of AI: offers orchestration/intervention guidance for those running sprints (abstract) | (b) AI planning/running/supporting group sessions: GenAI inside facilitated innovation sprints | (c) AI and group ideation/decisions: helps divergence, weak on empathic depth for convergence. - **Caveats:** Abstract only; qualitative; method details and participant counts unverified; practitioner-journal format. - **Screening check:** Likely meets criteria (71 teams using GenAI in collaborative design thinking; qualitative method and team count reported) — full method not verified. #### OA1070 — GenAI in a corporate Design Thinking workshop (case study) - **Citation:** Pena, T., Cunha, A. S., Cordeiro, J. V., & Victorino, G. (2026). Enhancing corporate innovation through generative AI: A case study on Human–AI collaboration. Social Sciences & Humanities Open, 13, 102867. https://doi.org/10.1016/j.ssaho.2026.102867 - **Type:** journal — peer-reviewed: yes (received 29 Sep 2025, revised 21 Apr 2026, accepted 1 May 2026; p. 1) - **Full text read:** yes (repository copy via http://hdl.handle.net/10362/203323, RUN NOVA; CC BY) - **Method:** Exploratory single case study, mixed methods, in-person corporate DT workshop with three 30-min exercises (trend identification, brainstorming, constructive feedback); three randomly assigned groups: "Humans" (no AI), "GenAI-Enhanced" (AI only after own work), "GenAI-Free Access" (optional AI), plus an AI-only benchmark (one non-expert researcher prompting ChatGPT); anonymous wisdom-of-the-crowd (WOC) votes (excluding own group) and a CEO "leadership" vote per exercise; pre/post 7-point Likert questionnaires, paired t-tests (SPSS 28) (§2.1–2.4, p. 3–4). - **n:** 15 employees from multiple departments (3 groups of 5); 14 completed pre and post questionnaires; median age 46; roles mostly directors/VPs (§2.2, §3, Table 4, Table 5; p. 4, 6, 8). - **Setting:** Customer-experience outsourcing company (sectors: banking, healthcare, retail), physical workshop; ChatGPT version 3.5 (§2.1, §2.3, p. 3–4). - **Findings:** 1. Ideation (brainstorming) votes: AI-only benchmark idea received the most WOC votes (6, reported as 40%); Humans 4, GenAI-Free Access 3, GenAI-Enhanced 2; CEO chose the GenAI-Free Access idea. Inspiration (trend) phase: GenAI-Free Access most WOC votes (5; reported as 33.3%), CEO chose GenAI-Enhanced (§3.3.1–3.3.2, Table 5, p. 6–8). 2. Pre→post: confidence in identifying trends 71.4% → 100% (levels 6–7); confidence in generating creative ideas 35.7% → 64.1% (levels 6–7), reported as statistically significant (p < 0.05, Fig. 5); belief that GenAI identifies user needs better than traditional methods 21.4% → 85.7% (levels 6–7) (§3.3.4, p. 9). Effect sizes not reported (§4.1, p. 11). 3. Authors: leadership votes and WOC diverged — leadership aligned with feasibility/strategic relevance, collective votes with novelty; "human or hybrid feedback frequently selected by leadership and WOC"; ethical-concern ratings largely unchanged post-workshop (§4, Table 7, p. 9–10). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: compares AI-use configurations within a facilitated corporate DT workshop | (c) AI and group ideation/decisions: AI-assisted and AI-only outputs often won peer votes, but leadership selections diverged; perception gains are self-reported. - **Caveats:** n = 15, one workshop, one group per condition (condition = group, confounded); "votes" from very few voters; no expert/blinded quality rating beyond peer and CEO; self-report pre/post with likely novelty effect (authors acknowledge, §4, §4.1); ChatGPT 3.5; some percentages (e.g., 64.1% of 14) not reproducible from counts — reported as stated. - **Screening check:** Meets criteria (empirical; groups of 5 employees working with ChatGPT in a DT workshop; method and n reported; English). #### R1655 — Human-genAI relationships in a group co-design workshop - **Citation:** Vartiainen, H., Liukkonen, P., & Tedre, M. (2024). Emerging human-technology relationships in a co-design process with generative AI. Thinking Skills and Creativity, 56, 101742. https://doi.org/10.1016/j.tsc.2024.101742 - **Type:** journal — peer-reviewed: yes - **Full text read:** yes — preprint version "The Mosaic of Human-AI Co-Creation: Emerging human-technology relationships in a co-design process with generative AI" (EdArXiv/OSF, https://doi.org/10.35542/osf.io/g5fb8); page numbers refer to the preprint PDF; published version not read - **Method:** research-creation approach; qualitative thematic analysis (Braun & Clarke) of video data from project presentations and joint reflections, with written reflections as supplementary data (pp. 5-6 of preprint) - **n:** 33 pre-service teachers in small groups of 3-4, across three identical four-hour workshops; 161.5 minutes of video (pp. 4-5 of preprint) - **Setting:** higher education course in Finland, fall 2023; pre-service teachers majoring/minoring in craft science; after a 2-h lecture, groups collaboratively designed a speculative biomimetic product for a dark, wet dystopian scenario using text-to-image genAI on one laptop per group (figure credits Midjourney) (pp. 4-6 of preprint) - **Findings:** 1. Four themes: AI as a tool in the brainstorming cycle; AI as a dialogue partner; AI as a shared object of activity in human-machine teaming; AI as an agent in the decision-making loop (pp. 7-10 of preprint). 2. A gap in interaction with genAI caused four groups to abandon some of their initial design ideas (p. 8 of preprint). 3. Group 1 experienced that the AI-generated shared object enabled extensive negotiation of meaning and made shared understanding and consensus-making faster (p. 9 of preprint); students felt genAI narrowed the design space and played a significant role in decision-making, though not seen as overpowering (p. 10 of preprint). - **Relevance:** (a) — | (b) — | (c) genAI in group co-design distributes decision-making, speeds consensus, and can steer or drop ideas - **Caveats:** qualitative, no comparison group; preprint read rather than published version; student (pre-service teacher) sample; text-to-image generator rather than LLM - **Screening check:** meets criteria (education exception) — generative AI takes part in a structured small-group co-design workshop and group-level design/decision processes are analysed; method and n reported; English. #### OA1993 — The AI penalty in deliberation (survey experiment, Germany) - **Citation:** Jungherr, A., & Rauchfleisch, A. (2025). Artificial intelligence in deliberation: The AI penalty and the emergence of a new deliberative divide. Government Information Quarterly, 42(4), 102079. https://doi.org/10.1016/j.giq.2025.102079 (preprint: arXiv:2503.07690) - **Type:** journal — peer-reviewed: yes (journal version); read here as arXiv preprint v1 (10 Mar 2025), page numbers refer to that preprint - **Full text read:** yes (https://arxiv.org/pdf/2503.07690, v1; the published GIQ version was not accessed and may differ) - **Method:** preregistered (OSF) vignette survey experiment; each respondent saw 6 of 12 deliberation tasks, 3 randomly attributed to AI and 3 to human facilitators; DVs: willingness to participate and a 5-item deliberative-quality index; multilevel models with varying intercepts for participants and tasks; IRB Univ. of Bamberg (Data & Methods, pp. 6–8). - **n:** 1,850 respondents (after removing 97 who failed ≥2 of 3 attention checks), Ipsos online panel with quotas representative of the German population 18+; ~11,000 task-level observations (461–464 per task); fielded 11–18 Feb 2025 (pp. 6, 8). - **Setting:** hypothetical descriptions of AI vs. human performance of deliberation tasks: recruitment/topic selection, summarisation, argument highlighting/structuring, opinion aggregation, fact-checking, disruption detection, tone & authenticity, gamification, translation, role-play empathy (p. 6). No actual AI system or group session. - **Findings:** 1. "AI-penalty": information that tasks were AI-performed lowered willingness to participate (b = −0.30, 95% CI [−0.34, −0.26], p < .001) and perceived deliberative quality (b = −0.29, 95% CI [−0.32, −0.26], p = .001) relative to human facilitation (p. 10). 2. Exploratory task-level analysis (not preregistered): respondents especially disliked AI for information and argument processing (summarisation, argument highlighting/structuring, opinion aggregation), while AI fact-checking was rated similar to human facilitators (pp. 8, 14–15; Fig. 1 p. 9). 3. The penalty depended on AI attitudes: positive interactions with perceived AI benefits (willingness b = 0.07; quality b = 0.10) and anthropomorphism (b = 0.10; b = 0.12), negative with AI risk perceptions (b = −0.13; b = −0.14), all p < .001; respondents with high benefit/low risk views showed "little to no AI-penalty" (pp. 13–14). - **Relevance:** (a) — | (b) public acceptance of AI performing facilitation/moderation tasks in deliberative sessions: participants prefer human facilitators | (c) — (perceived, not measured, decision quality) - **Caveats:** hypothetical vignettes, respondents never experienced AI facilitation (authors note contrast with Tessler et al. 2024, p. 15); German sample only; measures intention, not behaviour; page numbers from arXiv preprint, not the GIQ article. - **Screening check:** FAILS (strict reading): individual survey respondents rating hypothetical descriptions; no group interacting with generative AI, respondents are not facilitators, and no LLM is evaluated. Highly relevant as context on acceptance of AI facilitation — include only if attitudinal studies on AI-facilitated deliberation are admitted. #### S2-448 — PolicyPad: collaborative prototyping of LLM policies - **Citation:** Feng, K. J. K., Kuo, T.-S., Chen, Q. Z. (J.), Cheong, I., Holstein, K., & Zhang, A. X. (2026). PolicyPad: Collaborative Prototyping of LLM Policies. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26), Barcelona, Spain. ACM, 25 pages. https://doi.org/10.1145/3772318.3791689 (also arXiv:2509.19680v3) - **Type:** conference — peer-reviewed: yes (CHI '26 full paper; the acknowledgments thank the anonymous reviewers). The copy read is the arXiv v3 version (25 Feb 2026). - **Full text read:** yes (local PDF pdf3/S2-448.pdf = arXiv 2509.19680v3, https://arxiv.org/abs/2509.19680). The PDF has no printed page numbers, so pages below are "p. X of PDF". - **Method:** (1) Formative observational study by contextual inquiry over 15 weeks (Jan–Apr 2024), run with OpenAI: 19 virtual workshops of 60–90 min, each led by one facilitator from the AI lab or the research team, recorded, transcribed and coded (p. 4–5 of PDF, §3.1). (2) Evaluation study: group-based within-subjects comparison of PolicyPad vs a baseline "copilot-style" editor with the same policy-informed model, order and tasks counterbalanced in a 2×2 factorial design. Each session was 90 min: two 25-min prototyping tasks plus an exit interview. Analysis combined reflexive thematic analysis, an LLM-assisted plus human-annotated novelty analysis of the policy statements (initial Cohen's κ = 0.41, disagreements resolved by discussion), and cross-group Likert ratings compared with Wilcoxon signed-rank tests (pp. 9–13 of PDF, §6). - **n:** Formative: 9 clinical mental-health experts (E1–E9) across 19 workshops; attendance varied weekly because of scheduling conflicts (p. 4 of PDF, §3.1.1). Evaluation: 22 domain experts (10 mental health, 12 law) in 8 groups of 2–4 (median 3), i.e. 4 mental-health groups (MH1–MH4) and 4 legal groups (L1–L4). 5 of the mental-health experts had also taken part in the formative study. 4 of the 8 groups contained people who knew each other professionally (p. 11 of PDF, §6.1; Table 3, p. 25 of PDF). - **Setting:** Mostly US-based domain experts (two legal experts were based in Europe and Asia) drafted LLM behaviour policies (conversational tone; guardrails) in small real-time groups. The first author facilitated all evaluation sessions. PolicyPad is a collaborative TipTap editor. The policy-informed model is Llama 3.3 70B Instruct Turbo (policy given as system prompt). GPT-4o generates version titles. o4-mini suggests policy statements from edited responses and runs automated heuristic evaluation. Novelty screening used GPT-4.1 (pp. 7, 9, 12 of PDF, §5.3, §6.3.2). The session modality of the evaluation study (remote vs in-person) is not verified. - **Findings:** 1. Groups produced more novel policy statements with PolicyPad: "51.9% of the policy statements drafted in PolicyPad were considered novel, compared to 18.2% from our baseline… 4 times that of the baseline (40 vs. 10)" (p. 14 of PDF, §7.2.1). 2. Cross-group Likert ratings of the policies did not differ between conditions: "W = 6.0; p = 0.65; Msystem = Mbaseline = 4 for important considerations, W = 20.0; p = 0.74 … for agreement" (p. 14 of PDF, §7.2.1). 3. Group dynamics. Heuristics "built common ground", and spotlighting scenarios into the shared editor improved collaboration; in the baseline, experts improvised workarounds such as screen sharing. "100% of the policy suggestions were accepted in our studies" (p. 13 of PDF, §7.1.1–7.1.2). However, "participants tended to agree with others in their group and rarely challenged or pushed back directly on others' input" (p. 14 of PDF, §7.1.4). The authors also state that "the facilitator for a workshop holds a major role and can influence the results with their facilitation style" (p. 17 of PDF, §9). - **Relevance:** (a) — (the facilitator is a researcher; facilitators' own AI use is not studied) | (b) An LLM-integrated tool supported facilitated small-group expert workshops through rapid iteration, automated heuristic checks and AI-suggested policy statements | (c) Increased the novelty of group output (51.9% vs 18.2% novel statements); low within-group challenge; every AI suggestion was accepted, which suggests possible over-reliance. - **Caveats:** Small n (22 experts, 8 groups). The novelty judgement was LLM-assisted with human annotators from the research team (κ = 0.41 before discussion). The system condition had 10 extra demo minutes (40 vs 30 min per task, with 25 min of prototyping in both). The formative study was run with OpenAI. The task is policy drafting, not a generic workshop. The citation year in s2_decisions.tsv (2025) differs from the paper (CHI '26, 2026). - **Screening check:** Meets the criteria. It is empirical, with small groups (2–4) of humans collaborating in real time with LLM involvement in facilitated workshop sessions, and it reports method and n. English. ## 3. Generative AI and group ideation, creativity and team decisions #### OA1280 — AI-Augmented Brainwriting (LLM in group ideation + GPT-4 idea evaluation) - **Citation:** Shaer, O., Cooper, A., Mokryn, O., Kun, A. L., & Ben Shoshan, H. (2024). AI-Augmented Brainwriting: Investigating the use of LLMs in group ideation. Proceedings of the CHI Conference on Human Factors in Computing Systems (CHI '24), pp. 1–17. https://doi.org/10.1145/3613904.3642414 - **Type:** conference — peer-reviewed: yes (CHI full paper); read as arXiv version - **Full text read:** yes (arXiv preprint version: https://arxiv.org/pdf/2402.14978v2, 29 Feb 2024; page numbers below are arXiv PDF pages, not CHI proceedings pages; ACM PDF blocked) - **Method:** Field user study in a course: 70-minute Brainwriting session (individual idea writing on shared Conceptboard → team co-creation with GPT-3 → team selection/development) with self-ratings and reflections (post-session and end of semester); semantic clustering and Latent Personal Analysis of idea vocabularies; thematic analysis of prompts/reflections; post-hoc comparison of a GPT-4 idea-evaluation engine (29 successful runs) with 3 expert and 6 novice raters (Fleiss' kappa, Pearson correlation of rankings). No non-AI control condition (§4, §5.1, §5.3; p. 8–11, 17). - **n:** 16 undergraduate students (ages 18–23) in 5 teams of 3–4; 148 ideas (human 97, GPT-3 29, collaborative 22 per Table 1); raters: 3 experts (a 4th with partial ratings excluded), 6 novices (§4.1, Table 1, §5.1; p. 8–11). - **Setting:** Advanced undergraduate tangible interaction design course, February 2023 (Wellesley College affiliation); task: tangible UI ideas for people who work/study in mobile environments; OpenAI Playground GPT-3 for ideation; GPT-4 (June 2023, via API) for evaluation, not shown to teams (§3.1.2, §3.3.1, §4.1; p. 6–9). - **Findings:** 1. All 5 teams' final project ideas involved GPT-3: 3 combined one human and one GPT-3 idea, 1 combined multiple human and GPT-3 ideas, 1 was based solely on a GPT-3 idea (§5.2.2, Table 3, p. 13–14). 2. Reflections (n = 16): 50% (8/16) said GPT-3 offered a unique or expanded viewpoint; 44% (7/16) said it significantly assisted idea generation; 31% (5/16) said it was redundant and lacked creativity; 25% (4/16) reported prompt-crafting difficulties (§5.2.1, p. 12–13). GPT-3 ideas were more concrete/detailed (avg 20.9 vs 16.5 words) (§4.1, p. 10; §5.2.3, p. 15). 3. GPT-4 evaluation engine: moderate internal consistency (Fleiss' κ relevance 0.42, innovation 0.40, insightfulness 0.49); rated ideas higher than humans (means 4.19/3.72/3.68 vs experts 3.57/2.79/3.01); ranking correlations r = 0.556 (experts–GPT-4), 0.547 (novices–GPT-4), 0.602 (experts–novices); no team-chosen idea was rated low by GPT-4; largest disagreement: team 5 idea innovation 2.00 (experts) vs 4.93 (GPT-4) (§5.3.1–5.3.3, Table 4, p. 17–20). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: LLM integrated into a structured Brainwriting protocol; GPT-4 tested offline as a convergence/evaluation aid (an LLM doing a facilitation-like screening function) | (c) AI and group ideation/decisions: LLM ideas fed into all final team choices; LLM ratings moderately track human rankings but are inflated. - **Caveats:** Small single-class sample, no control group, novice GPT-3 users, single problem statement (§6.4, p. 23); GPT-4 evaluation post hoc; self-ratings heavily skewed (60% at maximum, p. 10). Text states "0 men"; arXiv version may differ from CHI version. - **Screening check:** Meets criteria (empirical; teams of 3–4 co-create with an LLM in a group ideation session plus offline LLM evaluation of ideas; method and n reported; English). #### OA1433 — GenAI in 3-person group brainstorming (IEICE 2025) - **Citation:** Kamada, K., Manakitrungrueng, N., & Yuizono, T. (2025/2026). Simply incorporating generative AI into groups is not enough: Exploring the effects of human-AI collaboration in groups on brainstorming. IEICE Transactions on Information and Systems, E109.D(5), 633–641 (advance publication 2025-11-28). https://doi.org/10.1587/transinf.2025DKP0002 - **Type:** journal — peer-reviewed: yes - **Full text read:** yes (https://www.jstage.jst.go.jp/article/transinf/advpub/0/advpub_2025DKP0002/_pdf — advance-publication version; page numbers below are the manuscript page numbers printed on that PDF, not the final 633–641 pagination) - **Method:** lab experiment (online), within-subjects, 3 conditions over 4 rounds with counterbalancing groups: (A) 3-person groups without GenAI, (B) 3-person groups with a shared GenAI agent, (C) individuals with GenAI; 15-min electronic brainstorming per topic on Slack; ideas rated by 2 independent human raters (7-point scales) and by embedding-based divergence/similarity; post-round questionnaires on self/other-perceived effort and psychological safety; IRB-approved (§3, p. 2–4). - **n:** 24 participants (19 male, 5 female; mean age 24.8, SD 2.3), split into two counterbalancing groups of 12; 2 evaluators; 4 rounds/topics (§3.1, Table 1–2, p. 2–3). Number of 3-person groups per round not stated explicitly. - **Setting:** Japanese-speaking students at JAIST (Japan), July–Aug 2025; topics: raising a town's café/cinema sales, park/library users; agent in Slack via GPT-4o mini API through "AI by Zapier", returning three ideas per "@ai" prompt, visible to all group members in condition B (§3.2–3.3, p. 3). - **Findings:** 1. Human-generated ideas per group fell with GenAI: condition A M = 17.4 (SD 7.2) vs. condition B M = 8.1 (SD 7.1); Mann-Whitney U = 54.0, Z = 2.31, p = .02, r = .58; human ideas were 39.7% (SD 0.22) of all ideas incl. GenAI in round 2 (§4.1, Table 3, p. 4). 2. No significant quality differences A vs. B on any measure (divergence p = .13; similarity p = .37; originality p = .60; elaboration p = .17; flexibility p = .46; appropriateness p = .96; feasibility p = .92); self-perceived effort marginally lower with GenAI (5.67 vs. 5.03; W = 24.5, Z = -1.81, p = .07, r = -.37); self–other effort correlation r = .61 (p = .002) without GenAI vs. r = .19 (p = .37) with GenAI; psychological safety unchanged (p = .50) (§4.1, Tables 3–4, p. 4–5). 3. GenAI-assisted groups vs. GenAI-assisted individuals: higher originality (3.94 vs. 3.19; p = .04, r = -.43), elaboration (4.22 vs. 3.41; p = .05, r = -.43) and flexibility (4.29 vs. 3.58; p = .02, r = -.51); fewer ideas per person, marginal (4.30 vs. 6.60; p = .06); human ideas were 26% (round 3) and 32% (round 4) of all ideas (§4.2, Table 5, p. 5–6). Authors recommend against unrestricted GenAI use in groups and suggest structured phases (human / GenAI / recombination) (§5.1, p. 7). - **Relevance:** (a) — | (b) shared GenAI agent inside a group brainstorming channel; implications for structuring AI phases in sessions | (c) GenAI halved human idea output in groups without improving quality; possible social loafing - **Caveats:** small n (24), student sample, text-only online Slack setting, short 15-min rounds, round/topic confounds handled only by counterbalancing; condition B vs. C comparison uses different rounds/topics; advance-publication pagination. - **Screening check:** meets criteria — empirical experiment, n reported, 3-person human groups interacting with an LLM agent. #### OA1366 — LLM bots in student group brainstorming (software engineering) - **Citation:** Farah, J. C., La Scala, J., Ingram, S., & Gillet, D. (2025). Supporting Brainstorming Activities with Bots in Software Engineering Education. IEEE/ACM International Workshop on Bots in Software Engineering (BotSE 2025), pp. 23–27. https://doi.org/10.1109/botse67031.2025.00013 - **Type:** conference (workshop short paper) — peer-reviewed: unclear (workshop proceedings; review process not stated in text) - **Full text read:** yes (https://infoscience.epfl.ch/bitstreams/77ea4453-786d-4e90-9933-a00734de7e7d/download, via https://infoscience.epfl.ch/handle/20.500.14299/246663.2) - **Method:** Within-subjects empirical case study; each group did 4 brainstorming tasks (~1 hour total), 2 with LLM bots and 2 without, randomly allocated to one of four treatment/control sequences; 3 idea rounds per task, then anonymous voting (3 votes each); post-task 7-point Likert items; Wilcoxon signed-rank tests with rank-biserial r, Kruskal-Wallis (§III, p. 2–3 of PDF). - **n:** 24 undergraduate students (14 data science, 10 software engineering) in 7 groups; 2–4 present per group (mean 3.43) on the day; 390 student ideas (§III, §IV-A, p. 2–3 of PDF). - **Setting:** College of Engineering and Architecture of Fribourg (HES-SO), October 2024, frontend development course project; collaborative brainstorming app on Graasp platform; two bots (HCIBot, SEBot) on GPT-4o (gpt-4o-2024-08-06, temperature 1), contributing one idea per three student ideas, tagged "AI Generated"; each group had a student "manager" controlling timing; experimenter present (§III-B, §III-C, p. 2–3 of PDF). - **Findings:** 1. Students proposed significantly fewer ideas when bots were present (W = 27.0, p = 0.00602, r_rb = −0.913) (§IV-A, p. 3 of PDF, Fig. 3). 2. 96.25% of votes went to student ideas; bots received only 10 votes in total; only one bot idea received the maximum votes across all tasks and groups (§IV-A, p. 3 of PDF). 3. With bots, lower sense of responsibility (W = 0.0, p = 0.0412, r_rb = −1.0) and authorship (W = 3.5, p = 0.0075, r_rb = −0.988) for selected ideas; no significant differences in perceived originality (p = 0.128) or usefulness (p = 0.823) (§IV-B, p. 4 of PDF). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: LLM bots as automated participants in a structured (rounds + voting) group brainstorm | (c) AI and group ideation/decisions: bots reduced human idea output and ownership of chosen ideas without raising perceived originality/usefulness; groups rarely chose bot ideas. - **Caveats:** Small n (7 groups, 24 students, 2 absent on the day); classroom setting; fixed task order; self-report Likert outcomes; very large effect sizes on small samples. Same research group and app as OA931 (La Scala et al.) — possible related samples, not verified. - **Screening check:** Meets criteria (empirical; student groups of 2–4 brainstorm with LLM bots contributing in shared session; method and n reported; English). #### OA1297 — ChatGPT and Me: group brainstorming with ChatGPT in a Teams chat - **Citation:** Kim, H. K., Roknaldin, A., Nayak, S. P., Zhang, X., Yang, M., Twyman, M., Hwang, A. H.-C., & Lu, S. (2024). ChatGPT and Me: Collaborative Creativity in a Group Brainstorming with Generative AI. American Society for Engineering Education (ASEE) Annual Conference 2024, Paper ID #43798. https://doi.org/10.18260/1-2--48457 - **Type:** conference — peer-reviewed: unclear (ASEE conference paper; review process not stated in text) - **Full text read:** yes (https://peer.asee.org/48457.pdf) - **Method:** Classroom quasi-experiment over two weeks: groups randomly assigned; half the groups (GA) had ChatGPT (bot "iBot") inside their Microsoft Teams group chat, the other half (GB) had no ChatGPT; each group submitted 10 product ideas/week; then merged groups (Gcomb) picked the top 5 most creative of 20 ideas and attributed their origin. Wilcoxon signed-rank test on share of ChatGPT-group ideas selected; regression for group/week effects; NLP (TF-IDF cosine distance, term-frequency, Jensen–Shannon divergence) (p. 4–9 of PDF). - **n:** 33 graduate students in 8 brainstorming groups of 3–5; 4 combined groups (9, 7, 9, 8 students) per week; 2 weeks (p. 4–5 of PDF, Table 2). - **Setting:** Graduate product engineering course, large private US university (western region), spring 2023; task: 10 COVID-prevention product ideas per week for given user scenarios; ChatGPT integrated in Microsoft Teams group chat (model version not verified) (p. 4 of PDF, Table 1, Fig. 1). - **Findings:** 1. Merged groups selected ChatGPT-group ideas more often than chance: E[RA] = 0.725 (≈3.625 of top 5 ideas from ChatGPT groups), Wilcoxon p = 0.012; week, group and share-of-GA-students not significant in regression (p = 0.444, 0.397, 0.411) (p. 6 of PDF). 2. ChatGPT-group ideas were longer (14.92 vs 10.5 words after preprocessing, significant at 0.001 level) and used more unique words (573 vs 376); Jensen–Shannon divergence 0.4094 — authors propose descriptiveness as one possible reason for higher perceived creativity (p. 7 of PDF). 3. No evidence of homogenisation: pairwise TF-IDF cosine distances between groups' idea sets 0.8–0.92 in both conditions; t-tests p = 0.9624 (week 1) and 0.3375 (week 2) (p. 9 of PDF). Chat logs show GA members "relied heavily on ChatGPT to generate their initial ideas" but modified them (p. 5–6 of PDF). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: ChatGPT embedded as a reactive participant in a group chat brainstorm | (c) AI and group ideation/decisions: ChatGPT-assisted group ideas were chosen as more creative by peers; no loss of between-group idea diversity. - **Caveats:** Small n (8 groups, 2 weeks); "creativity" = peer selection by mixed groups (possible bias, though tested); raters partly members of originating groups; no expert rating; ChatGPT model/version not stated; peer-review status unclear. - **Screening check:** Meets criteria (empirical; student groups of 3–5 interact with ChatGPT in shared group chat; method and n reported; English). #### OA1461 — ChatGPT in asynchronous student group brainstorming (CSCL) - **Citation:** Kim, H. K., Nayak, S., Roknaldin, A., Zhang, X., Twyman, M., & Lu, S. (2024). Exploring the impact of ChatGPT on student interactions in computer-supported collaborative learning. arXiv:2403.07082 [cs.CY]. https://doi.org/10.48550/arXiv.2403.07082 - **Type:** preprint (2-page short paper; carries an AAAI 2024 copyright line, but the venue is not named) — peer-reviewed: no (preprint) / unclear - **Full text read:** yes (https://arxiv.org/pdf/2403.07082) - **Method:** field experiment in a course: cohorts randomly assigned to have (Type A) or not have (Type B) ChatGPT as a taggable question-answering agent in asynchronous group chat; two expert annotators coded messages with Soller's Collaborative Learning Conversation Skills Taxonomy (Request/Inform/Motivate = "active learning"), agreement reported as Cohen's kappa 93.2%; repeated ANOVA (ChatGPT × week) on proportion of active-learning messages, Tukey post-hoc (pp. 1–2 of PDF). - **n:** 12 cohorts of 4–5 students (6 with ChatGPT, 6 without); 57 students (29 Type A, 28 Type B); two weekly sessions of 3 days each (p. 1, Table 5 p. 3 of PDF). - **Setting:** graduate product-engineering course (University of Southern California), spring 2023; asynchronous online group brainstorming of product design ideas for given customer requirements; ChatGPT via OpenAI API (model version not stated) (p. 1 of PDF). - **Findings:** 1. No significant difference in the overall share of active-learning messages between cohorts with and without ChatGPT (ChatGPT F = 0.202, df 1/20; Table 8, p. 3 of PDF). 2. Student-to-student messages: cohorts without ChatGPT had a 22.6 percentage-point higher ratio of active learning (ANOVA F = 4.909, p < .05; Tukey difference 0.226, adj. p = 0.037) (Tables 1–2, p. 2 of PDF). 3. Within ChatGPT cohorts, messages directed to ChatGPT had a 53.57 percentage-point higher level of active learning than messages to fellow students (F = 24.765, p < .001; adj. p = 0.000); authors conclude active learning became "more bilateral between students and ChatGPT" (Tables 3–4, p. 2 of PDF). - **Relevance:** (a) — | (b) LLM agent added to asynchronous group brainstorming channels | (c) ChatGPT shifted substantive interaction from peer-to-peer to student-to-AI; idea outcomes not measured - **Caveats:** very short preprint; small number of units (12 cohorts); asynchronous text chat; outcome is interaction coding, not idea quality; novelty effect not controlled (acknowledged p. 3); message tables contain duplicated cohort labels ("Cohort #4" twice, Tables 6–7). - **Screening check:** meets criteria — empirical, n reported, groups of 4–5 humans interacting with an LLM. #### C605 — AI-deation: AI collaborator and LLM idea editor in group ideation - **Citation:** La Scala, J., Bartłomiejczyk, N., Gillet, D., & Holzer, A. (2025). Fostering Innovation with Generative AI: A Study on Human-AI Collaborative Ideation and User Anonymity. Proceedings of the 58th Hawaii International Conference on System Sciences (HICSS), pp. 342–351 (first page 342 as printed; last page not verified). https://doi.org/10.24251/hicss.2025.041 / https://hdl.handle.net/10125/108877 - **Type:** conference — peer-reviewed: yes - **Full text read:** yes (https://scholarspace.manoa.hawaii.edu/bitstreams/fa924f48-1b21-44f1-811e-79e4285d23ec/download) - **Method:** Randomized between-subjects classroom experiment (two conditions: human-only vs. humans + AI collaborator), post-questionnaire (S-UEQ, 13 seven-point Likert items, five open questions), novelty ratings by participants, innovation ratings by two authors on the Shaer et al. (2024) scale; thematic analysis of open answers (§4–6). - **n:** 32 students in course, 29 consented (9 male / 23 female among the 32); 13 participants in control, 16 in treatment; course groups: 11 groups of 3 and 1 group of 2 (p. 345, §4; p. 346, §5). 201 ideas total (77 control human ideas; 76 treatment human ideas + 48 AI ideas) (p. 346, §5). 25 answered open questions (p. 348, §6). - **Setting:** Master's students in General Management, information system design course (SDG-related group project), one class session; web app AI-deation (Graasp platform); three rounds of three minutes of idea generation, then rating; AI collaborator and idea editor both GPT-3.5 Turbo; AI collaborator added two ideas per round in treatment; idea editor rewrote all submitted ideas to strengthen anonymity (pp. 344–346, §3–4). - **Findings:** 1. AI-generated ideas were rated more novel than participants' ideas by participants (Mann-Whitney U = 4463, p = .012) and more innovative by raters (U = 1629, p < .001), but participants' ideas in the AI condition were not more novel (U = 3501.5, p = .983) or more innovative (U = 2803, p = .8) than in the human-only condition (pp. 346–348, §5.2). 2. The AI collaborator did not increase the number of ideas per participant (Welch t(24.6) = 1.33, p = .2); authors conclude it "did not act as a catalyst for increasing idea generation" (p. 348, §5.2). 3. Participants significantly agreed the LLM idea editor hid their identity (t(28) = 1.99, p = .028), that anonymity made them more comfortable (t(27) = 3.07, p = .002) and that the editor was helpful (t(28) = 2.35, p = .013); open answers also reported the editor sometimes changed meaning or made ideas too general (pp. 346, 348–349, §5.1, §6). S-UEQ mean 0.2 = neutral usability (p. 346, §5). - **Relevance:** (a) — | (b) LLM as an embedded ideation tool (anonymising editor, AI co-participant) in a structured group ideation session | (c) AI ideas more novel/innovative, but no spillover to human ideas' novelty, innovation or quantity - **Caveats:** Small student sample (n = 29), single 9-minute ideation session, non-native English speakers; no control over participants using other AI tools (p. 346, §4); innovation rated by two authors (r = .79 after consensus discussion, p. 347). - **Screening check:** Meets criteria — empirical randomized experiment with small groups ideating together with an LLM (AI collaborator + editor); method and n reported; English. #### C60 — GraftMind: AI-mediated idea sharing in group ideation - **Citation:** Li, J., Fang, Y., He, X., & Fan, M. (2026). GraftMind: Facilitating Group Ideation with AI-Mediated Idea Sharing. Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26), Barcelona, pp. 1–18. https://doi.org/10.1145/3772318.3791388 - **Type:** conference — peer-reviewed: yes - **Full text read:** yes (author copy https://www.mingmingfan.com/papers/CHI26-GraftMind.pdf) - **Method:** Between-subjects user study (GraftMind vs. control shared digital whiteboard for conventional group brainstorming), plus exploratory phase, questionnaires (synergy, evaluation apprehension scales), and semi-structured interviews with thematic analysis (p. 2–3, p. 8–10 of PDF, §5). - **n:** 60 participants (28 undergraduates, 32 graduate students), 30 per condition, 10 groups of 3 per condition = 20 groups; one 20-minute ideation session per group (p. 2–3, p. 9 of PDF). - **Setting:** University students (HKUST Guangzhou) brainstorming "What skills should students focus on cultivating for the future with the development of AI?" in private workspaces; a proactive AI mediator (GPT-4.1) abstracts others' ideas into higher-level directions and delivers persistence, flexibility and distinctiveness assistance (p. 1, p. 6, p. 9 of PDF). - **Findings:** 1. GraftMind produced significantly more ideas than the control (t(58) = 2.342, p = 0.015) and more distinct idea clusters, an originality proxy computed with BGE-M3 embedding clustering (t(58) = 2.177, p = 0.022) (p. 10–11 of PDF, §6.1.1). 2. Evaluation apprehension was significantly lower with GraftMind (one-tailed Mann–Whitney U = 141.5, p < 0.001) (p. 13 of PDF, §6.3.1). 3. Perceived synergy was not significantly lower than in conventional group brainstorming (t(58) = −1.02, p = 0.156), so private ideation with AI-relayed ideas kept synergy comparable (p. 12 of PDF, §6.2.1). The interviews produced 76 codes (p. 10 of PDF). - **Relevance:** (a) — | (b) AI mediator as a real-time facilitation layer that routes ideas between group members | (c) more ideas and more distinct idea clusters, lower evaluation apprehension, synergy unchanged - **Caveats:** Student sample. The authors say condition assignment was "not fully randomized" because the control condition ran later (p. 3 of PDF). Tests use df = 58 (participant level) although ideas are described per group. Originality is an embedding-cluster proxy, not human-rated. - **Screening check:** Meets criteria (empirical, groups of 3 humans interacting via an LLM mediator, method and n reported, English). #### C18 — Controlling AI agent participation in group conversations (Koala) - **Citation:** Houde, S., Brimijoin, K., Muller, M., Ross, S. I., Silva Moran, D. A., Gonzalez, G. E., Kunde, S., Foreman, M. A., & Weisz, J. D. (2025). Controlling AI Agent Participation in Group Conversations: A Human-Centered Approach. In Proceedings of the 30th International Conference on Intelligent User Interfaces (IUI '25), pp. 390–408. https://doi.org/10.1145/3708359.3712089 - **Type:** conference (full paper) — peer-reviewed: yes - **Full text read:** yes (arXiv version, https://arxiv.org/pdf/2501.17258v1) - **Method:** Two mixed-methods user studies. Study 1: within-group comparison of three conditions in timed Slack brainstorms (no AI / reactive AI / proactive AI), idea coding, preference survey, reflexive thematic analysis. Study 2: re-recruited groups used improved Koala II with a control-settings panel plus design probes and semi-structured discussion; outputs a taxonomy of controllable agent behaviors (§4, §6, p. 7–17 of PDF). - **n:** Study 1: 18 IBM employees in 6 groups of 3 (6 sessions); Study 2: 14 of the 18 re-recruited, same 6 groups (§4.1.2 p. 8; §6.1.1 p. 14). - **Setting:** Virtual 3-minute brainstorms in Slack on workplace topics (e.g., improving hybrid meetings), each session run by a human moderator; LLM agent "Koala" as group participant, driven by Llama 2 (Study 1) and Llama 3 (Koala II, Study 2) (§3, p. 6; §5, p. 13). - **Findings:** 1. 73% of all ideas produced in the AI conditions were contributed by Koala, and Koala's ideas made up 33% of the groups' selected top ideas; participants preferred having either Koala variant over no AI (§4.2.1, p. 9). 2. 13 of 18 participants (72.2%) preferred the reactive over the proactive variant; the proactive agent was described as distracting, intrusive and dominating ("Koala dominated the conversation"), with one participant noting the team generated fewer ideas because of interruptions (§4.2.2, p. 11). 3. With Koala II and a settings panel, no group chose to switch to the reactive variant; participants rated the value of controls at mean 4.46 (SD = 0.66) out of 5 (§6.2.2, p. 15–16). - **Relevance:** (a) — | (b) design of when/how an LLM agent should contribute to a moderated group session; user control over proactivity | (c) LLM agent dominated idea volume (73%) and share of top ideas (33%); risk of crowding out human contribution - **Caveats:** Small n; single company (IBM) sample; very short 3-min brainstorms; descriptive statistics, no significance tests on idea counts; human moderator ran sessions (AI was a participant, not facilitator). - **Screening check:** Meets criteria — empirical, groups of 3 humans interacting with an LLM agent in brainstorming, method and n reported. #### S2-567 — Group Brainstorming with an AI Agent (Koala) - **Citation:** Muller, M., Houde, S., Gonzalez, G., Brimijoin, K., Ross, S. I., Silva Moran, D. A., & Weisz, J. D. (2024). Group Brainstorming with an AI Agent: Creating and Selecting Ideas. International Conference on Computational Creativity (ICCC 2024), paper 18. https://computationalcreativity.net/iccc24/papers/ICCC24_paper_18.pdf (venue/edition not printed in the PDF text; taken from URL; page range not verified) - **Type:** conference — peer-reviewed: yes (ICCC; not stated in the PDF itself) - **Full text read:** yes (https://computationalcreativity.net/iccc24/papers/ICCC24_paper_18.pdf; local pdf3/S2-567.txt) - **Method:** counterbalanced within-group study (3 topics × 3 sessions; baseline no-AI, Reactive AI, Proactive AI); this paper analyses transcripts of the two AI sessions only: 3 authors coded each idea's action sequence (idea/plus/minus/refine/final), disagreements resolved by discussion; chi-square and F tests (pp. 3–4). Survey data mentioned but not analysed here. - **n:** 18 participants (10 women, 7 men, 1 prefer not to answer; 3 countries; design, user research, videography, software engineering) in 6 groups of 3; each group 3 sessions (1 baseline, 2 with Koala); 171 ideas and 36 final ideas analysed from the AI sessions (pp. 1, 3, 5). - **Setting:** IBM employees brainstorming in Slack; Koala = multi-party conversational agent on Llama 2, labelled as AI ("APP" suffix), as "fourth participant"; Reactive (only when asked) vs. Proactive (decides itself when to contribute) versions; topics: hybrid meetings, HR chatbot engagement, conference give-aways; each session ends with the group selecting 3 "final" ideas for an imagined client; $25 compensation for one hour (p. 3). - **Findings:** 1. Survey data (reported in one sentence, no numbers beyond this): "100% of participants preferred AI-supported brainstorming to the no AI condition, and ... the AI conditions led to significantly more ideas" (p. 3); test statistics not reported. 2. AI proposed 124 ideas (73% of all 171), of which 84 (68%) were never discussed by anyone; humans proposed 47 (27%). Humans chose AI-initiated ideas as final in 33% of cases; human-proposed ideas had a "two-to-one advantage" (pp. 4–5, 8). Human vs. AI action distributions differed, χ²(4) = 118.861, p < 0.001 (adjusted χ²(4) = 72.476) (p. 4). 3. 28 of 36 final ideas (78%) involved both a human and an AI action vs. 39 nonfinal (29%); human–AI co-involvement "more than doubled" the likelihood of becoming final, χ²(1) = 8.847, p < 0.003; 89% of final vs. 33% of nonfinal ideas had supportive (plus) actions, χ²(1) = 36.480, p < 0.001; final ideas were decided early (69% by step 7), F(1,169) = 134.052, p < 0.001 (pp. 5–7). - **Relevance:** (a) — | (b) LLM agent participating (reactive/proactive) in live group brainstorming sessions | (c) AI generated most ideas but most were ignored; final selections favoured human ideas and human–AI co-developed ideas - **Caveats:** Small n (6 groups); participants are colleagues at the authors' company; Reactive vs. Proactive conditions pooled in these analyses (no condition comparison reported); baseline vs. AI comparison only as a one-line survey claim without statistics; idea-level chi-square tests treat ideas as independent; overlaps with S2-569 (same IBM team, same Llama 2 backend family). - **Screening check:** Meets criteria — empirical; groups of 3 humans interacting with an LLM agent; method and n reported; English. #### C52 — CreAItive Collaboration? Misjudgment of AI creativity in dyadic ideation - **Citation:** Bangerl, M. M., Disch, L., David, T., & Pammer-Schindler, V. (2025). CreAItive Collaboration? Users' Misjudgment of AI-Creativity Affects Their Collaborative Performance. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI '25), pp. 1–17. https://doi.org/10.1145/3706598.3713886 - **Type:** conference (full paper) — peer-reviewed: yes - **Full text read:** yes (TU Graz repository PDF, https://tugraz.elsevierpure.com/ws/portalfiles/portal/97638468/3706598.3713886.pdf) - **Method:** Between-subjects comparative experiment: human dyads randomly assigned to AI condition (off-the-shelf ChatGPT-4 available) vs. control (no AI) performing the Alternate Uses Test (AUT) three times on-site, then an unaided immediate post-test in dyads and an unaided individual online retention post-test after 3 weeks; AUT scored for fluency, flexibility, originality, elaboration; Mann-Whitney U tests; coding of prompts and acceptance/rejection of AI output; questionnaires (§4, p. 4–7 of PDF). - **n:** 52 dyads / 104 participants (26 dyads per condition); aged 18–35 (mean 21.55), 77.88% male; 71.43% knew their partner (§4.3, p. 6). - **Setting:** Students (German-speaking) in a lab, partners seated opposite each other on separate laptops with a shared display; 4 minutes per AUT item; ChatGPT-4 (§4.2, p. 5–6). - **Findings:** 1. Main test: no significant differences between AI and control dyads in fluency, originality or flexibility, but AI dyads elaborated significantly less (Mann-Whitney U, statistic 924.0, p = 0.005); the elaboration deficit was replicated in the unaided immediate post-test (statistic 970.0, p = 0.012); no significant differences in the 3-week individual retention test (§5.1.1–5.1.3, p. 8–9). 2. Dyads accepted only 32.91% of ChatGPT-4-generated alternate uses (40.90% excluding duplicates/no-time), i.e. rejected 59.10%; entered answers in the AI group were ~50% AI-generated and 50% human (§5.2, p. 11). 3. Acceptance rate of AI answers correlated positively with fluency (Spearman 0.424, p = 0.002); authors interpret the selective rejection as misjudgment of AI creativity that reduced performance (§5.2, p. 11; Abstract). - **Relevance:** (a) — | (b) — | (c) GenAI access did not improve dyadic ideation quantity/originality, reduced elaboration, and groups under-used/misjudged AI ideas - **Caveats:** Student sample, mostly male; standardized AUT task rather than real workshop ideation; some statistical symbols garbled in PDF text extraction (test statistic letters not verified; values as reported). - **Screening check:** Meets criteria — randomized experiment with 2-person groups ideating jointly with generative AI (ChatGPT-4), method and n reported. #### C566 — Groups and AI: access to ChatGPT in co-located creative teamwork - **Citation:** Preuschoff, P. M., Zaghal, R., Nowak, O., & Borchers, J. (2025). Groups and AI: How Access to Generative AI Influences Creative Teamwork. In Mensch und Computer 2025 (MuC '25), pp. 407–421. ACM. https://doi.org/10.1145/3743049.3743073 - **Type:** conference — peer-reviewed: yes - **Full text read:** yes (https://hci.rwth-aachen.de/publications/preuschoff2025b.pdf; CC BY-NC-ND; page numbers = proceedings pages printed on the PDF) - **Method:** Study 1: within-group lab experiment, groups of three did two ~20-min creative marketing tasks (product name, slogan, 2–3 selling points, pen-and-paper sketch), once with ChatGPT-4o and once with Google search only, counterbalanced across 6 trials; questionnaires (Team Effectiveness Questionnaire items, Creative Product Semantic Scale, NASA-TLX), video observation with reflexive thematic analysis by one coder. Study 2: online study in which externals, blind to AI use, rated and ranked all results (§3, pp. 409–410). - **n:** Study 1: 18 participants in 6 groups of three (aged 20–29, mostly computer science/STEM, all prior AI users); Study 2: 20 external raters (§3.1–3.2, pp. 409–410; §8, p. 413). - **Setting:** RWTH Aachen university lab; co-located group at a table sharing one 55" screen with keyboard/mouse showing ChatGPT-4o or Google; fictitious products ("Glow-in-the-dark Umbrella", "Smart Plant Pot") (§3.1, p. 409). - **Findings:** 1. Only significant self-report effect: lower cognitive effort with AI (Z = −3.18, p ≤ 0.01; AI mean −0.56, SD 0.78 vs. no-AI 0.33, SD 0.84); self-rated creativity of results did not differ; 7 preferred the AI campaign, 4 no-AI, 7 no preference; most felt the group was more creative without AI (only 4 felt AI results were more creative); AI role perceived as guide (12), teammate (5), leader (1) (§4.1, p. 410). 2. External raters: effect of AI on "organized" (F1,57 = 5.43, p ≤ 0.05) and "well-defined/focus" (F1,57 = 4.38, p ≤ 0.05), with campaign interactions (post-hoc effects only for Smart Plant Pot); AI results "rated and ranked slightly better" (§4.1, p. 410; Fig. 3, p. 411). 3. Observed group process (descriptive frequencies, one coder): without AI, "Discussing Product Details" was four times more common, "Proposing a New Idea" nearly three times, "Critical Discussion" and "Concept Clarification" twice as often; with AI, groups often prompted ChatGPT 10–20 s into a session and "Parallel Working" was about twice as frequent (§4.2.1–4.2.2, pp. 410–412). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: implication that groups should ideate before consulting GenAI and that AI access reduces idea exchange (§5, pp. 412–413) | (c) AI and group ideation/decisions: AI access slightly improved externally rated output structure but reduced discussion, critical exchange and new-idea proposals in co-located groups - **Caveats:** Very small sample (6 groups, 18 people), young technical student sample, single coder for observations with frequency counts not tested statistically, non-expert external raters, short tasks. - **Screening check:** Meets criteria — empirical within-group experiment with co-located groups of 3 humans using ChatGPT-4o in a group ideation task, method and n reported, English. #### OA735 — Group ideation with an LLM in a virtual canvas (Collaborative Canvas) - **Citation:** He, J., Houde, S., Gonzalez, G. E., Moran, D. A. S., Ross, S. I., Muller, M., & Weisz, J. D. (2024). AI and the Future of Collaborative Work: Group Ideation with an LLM in a Virtual Canvas. Proceedings of the 3rd Annual Meeting of the Symposium on Human-Computer Interaction for Work (CHIWORK '24), pp. 1–14. https://doi.org/10.1145/3663384.3663398 - **Type:** conference — peer-reviewed: yes (not verified beyond venue) - **Full text read:** abstract only (ACM open access listed but blocked by ACM bot wall; no other copy found) - **Method:** user study with a technology probe (Collaborative Canvas, virtual canvas with integrated generative AI) (abstract); session format (individual vs. group) not verified - **n:** 17 professionals experienced with virtual group ideation workshops (abstract) - **Setting:** remote team ideation in a shared digital workspace; LLM model not verified - **Findings:** 1. Participants "saw value in using generative AI to assist with group facilitation and to augment perspectives and ideas" (abstract). 2. They "worried about losing human perspectives and critical thinking, as well as reputational harms resulting from harmful AI outputs" (abstract). 3. Identified needs: transparency of content ownership, private digital spaces, specialised AI capabilities (abstract). - **Relevance:** (a) experienced workshop practitioners on GenAI for facilitation | (b) LLM integrated into a virtual ideation canvas | (c) perceived risk to human perspectives/critical thinking - **Caveats:** abstract only; unclear whether participants ideated together in groups or individually with the probe; IBM authors/probe. - **Screening check:** Meets on the "facilitators/meeting professionals as respondents" route (17 professionals experienced in virtual group ideation workshops, GenAI probe, n reported); group-interaction format to confirm in full text. #### S2-569 — Collaborative Canvas: LLM Use in Group Ideation - **Citation:** Gonzalez, G. E., Silva Moran, D. A., Houde, S., He, J., Ross, S. I., Muller, M., Kunde, S., & Weisz, J. D. (2024). Collaborative Canvas: A Tool for Exploring LLM Use in Group Ideation Tasks. Joint Proceedings of the ACM IUI Workshops 2024 (Greenville, SC, USA, March 18–21, 2024), CEUR Workshop Proceedings Vol-3660, paper 4. https://ceur-ws.org/Vol-3660/paper4.pdf (volume/paper number from URL; not printed in text) - **Type:** workshop paper — peer-reviewed: unclear (CEUR workshop proceedings) - **Full text read:** yes (https://ceur-ws.org/Vol-3660/paper4.pdf; local pdf3/S2-569.txt) - **Method:** system description plus "preliminary evaluation": hour-long observed sessions with a demonstration, a short ideation task and joint clustering; findings reported as participant feedback bullet points; data-collection instrument and analysis method not described (§4, pp. 5–6). - **n:** 8 design professionals who facilitate design-thinking workshops as part of their work, in groups of 2–3 (§4, p. 5). Number of sessions not stated (not verified). - **Setting:** IBM (authors' organization); Collaborative Canvas — shared sticky-note canvas (Svelte/FastAPI) calling Llama 2 via internal IBM watsonx.ai; LLM can generate notes directly on the shared canvas or in a private scratchpad, with progressive refinement; task: generate as many ideas as possible on benefits of using LLMs in design-thinking sessions, then cluster by theme (§3.1, p. 4; §4, p. 5). - **Findings:** 1. Most participants preferred generating sticky notes in the personal scratchpad rather than directly on the canvas, to review and filter AI output before making it visible to the group (§4, pp. 5–6). 2. Facilitators felt generative capabilities could add value to design-thinking workshops by contributing novel ideas to get groups started and to get groups "unstuck" during lulls (§4, p. 6). 3. Facilitators envisioned AI support for their facilitator role beyond idea generation: clustering notes, keeping a workshop on schedule, summarising contributed ideas, managing phases (§4, p. 6; §5.2, p. 7). Participants found it "extremely important" to distinguish AI- from human-generated content (§5.1, p. 7). No quantitative results. - **Relevance:** (a) professional workshop facilitators' use of and wishes for LLM support in design-thinking sessions | (b) LLM as a tool inside a shared ideation canvas; envisioned facilitation functions | (c) AI notes used verbatim, modified, or as inspiration; no quality measure - **Caveats:** Very small, internal sample; preliminary, anecdotal feedback without described analysis or numbers; inference latency sometimes high and frustrating (§5.1, p. 7); same IBM team as S2-567. - **Screening check:** Meets criteria (weakly) — empirical preliminary evaluation with groups of 2–3 humans using an LLM together, respondents are workshop facilitators; method only loosely described but n reported; English. #### C15 — How creative design teams make sense of contrasting AI personas - **Citation:** Ma, K., Won, D., Zeng, Y., Chung, J., & Goucher-Lambert, K. (2026). How Creative Design Teams Make Sense of Contrasting AI Personas. ACM conference proceedings (2026), pp. 1633–1640. https://doi.org/10.1145/3803784.3816826 (venue name not verified; 8-page paper, reported as "preliminary findings") - **Type:** conference (short paper, per page range) — peer-reviewed: unclear (likely yes; not verified) - **Full text read:** abstract only (ACM DL blocked; no preprint found on arXiv or author pages) - **Method:** Counterbalanced within-subject study with two theory-grounded AI collaborator personas (control-oriented vs. autonomy-supportive) used as contrasting probes; mixed-methods analysis (Abstract). - **n:** 16 human design dyads (32 participants implied; not verified) (Abstract). - **Setting:** Open-ended design ideation tasks on a custom digital whiteboard with an AI collaborator; AI model not stated in abstract (not verified). - **Findings:** 1. Two emergent team orientations: persona-sensitive teams (7/16) treated the control-oriented persona as intrusive or adversarial, while persona-neutral teams (9/16) engaged both personas in a tool-like manner (Abstract). 2. Teams do not respond to AI personas uniformly; existing collaboration norms appear to shape responses (Abstract; described as preliminary). - **Relevance:** (a) — | (b) AI collaborator behaviour/persona in co-located team ideation sessions (relevant to designing an AI co-facilitator's social stance) | (c) team reactions to AI social behaviour during ideation; no outcome-quality data in abstract - **Caveats:** Abstract only; preliminary, short paper; small n; AI model and outcome measures unknown. - **Screening check:** Likely meets criteria (empirical, dyads of humans ideating with an AI collaborator, method and n reported in abstract); generative-AI/LLM involvement is not explicitly stated in the abstract — not verified. #### C446 — The Cybernetic Teammate (P&G field experiment) - **Citation:** Dell'Acqua, F., Ayoubi, C., Lifshitz, H., Sadun, R., Mollick, E., Mollick, L., Han, Y., Goldman, J., Nair, H., Taub, S., & Lakhani, K. (2025). The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise. NBER Working Paper No. 33641. https://doi.org/10.3386/w33641 - **Type:** preprint (working paper) — peer-reviewed: no (the paper states NBER working papers "have not been peer-reviewed", p. 1) - **Full text read:** yes (https://www.nber.org/system/files/working_papers/w33641/w33641.pdf) - **Method:** Pre-registered (AEA RCT Registry AEARCTR-0013603) randomised 2×2 field experiment (individual vs. two-person cross-functional team × with vs. without GenAI), stratified by business unit and geography, within a one-day virtual product-development workshop; solutions rated 1–10 by blind expert evaluators (>3 ratings per solution on average); OLS with robust SEs (§3–4, pp. 8–14). - **n:** 811 workshop participants, 776 randomised professionals (Commercial and R&D) at Procter & Gamble; teams = 1 Commercial + 1 R&D member; main quality regressions report 550 observations (solutions) (§3.2, p. 9; regression tables, appendix). Per-condition counts not verified. - **Setting:** P&G business units (Baby Care, Feminine Care, Grooming, Oral Care) in Europe and the Americas, May–July 2024; real new-product/innovation challenges; teams collaborated remotely via Microsoft Teams with one member sharing screen; GenAI tool built on GPT-4 via Microsoft Azure (GPT-4o in July session), plus 1-hour prompting training for AI conditions (§3.1–3.2, pp. 8–10). - **Findings:** 1. Quality vs. individuals without AI: teams without AI +0.24 SD (p < 0.05); individuals with AI +0.37 SD (p < 0.01); teams with AI +0.39 SD (p < 0.01); teams with AI not significantly different from individuals with AI (§5.1, p. 14; §7, p. 20). AI reduced time spent by 16.4% (individuals) and 12.7% (teams) (§5.1, p. 15). 2. Teams with AI were 9.2 percentage points more likely to produce a top-10% solution vs. control mean of 5.8% (individuals with AI: small, non-significant effect) (§6.1, p. 17). Without AI, team solutions were bimodal on the technical–commercial scale (bimodality coefficient 0.564) vs. unimodal with AI (0.482), interpreted as reduced dominance of one member (§6.3, p. 18). 3. Positive emotions rose by 0.457 SD (individuals with AI) and 0.635 SD (teams with AI) (p < 0.01), negative emotions fell by −0.233 and −0.235 SD (p < 0.05); AI-enabled participants were 9.2 pp less likely to expect a top-10% result (p < 0.05); AI-aided solutions showed higher semantic similarity to each other than human-only solutions (§5.3, p. 16; §6.2, p. 18; §6.4, p. 19). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: AI as tool used by two-person teams in a one-day innovation workshop (not as facilitator) | (c) AI and group ideation/decisions: strong causal evidence on quality, top-decile ideas, balance of perspectives within pairs, and homogenisation of ideas - **Caveats:** Working paper, not peer-reviewed; teams are dyads only; one-day virtual task; AI used as a shared tool rather than as a facilitator; P&G funded the research institute and one author consulted for P&G (p. 1). - **Screening check:** Meets criteria — empirical pre-registered field experiment, 2-person human teams working with GPT-4 in a group ideation/innovation task, method and n reported, English. #### OA1752 — Information elaboration and AI-driven diversity in team ideation (AOM 2025) - **Citation:** Wang, Z., Cheng, X., Zhu, J. N. Y., Xie, W., & Ma, Y. (2025). From diversity to novelty: How information elaboration unlocks the creative power of AI in teams. Academy of Management Proceedings, 2025(1), 21657 (abstract). https://doi.org/10.5465/amproc.2025.21657abstract - **Type:** conference abstract (AOM Annual Meeting proceedings abstract) — peer-reviewed: unclear (AOM proceedings abstracts are reviewed submissions; full paper not published there) - **Full text read:** abstract only (only the proceedings abstract exists at this DOI; retrieved via OpenAlex/Crossref record) - **Method:** "five preregistered, randomized controlled field experiments" analysed with a moderated-mediation model (AI → idea diversity → novelty, moderated by team information elaboration) (abstract). Details not verified. - **n:** 528 individuals across 176 teams (abstract); team size not stated (528/176 = 3 on average — derived, not stated). - **Setting:** ideation phase of new product development (NPD) in organisational teams; generative AI tool not specified (abstract). - **Findings:** 1. "Generative AI significantly enhances idea diversity, [but] diversity alone does not directly translate into novelty" (abstract). 2. The effect of AI-driven diversity on novelty "depends on the level of information elaboration within the team"; teams that analyse and integrate diverse perspectives more deeply leverage AI-enhanced diversity better (abstract). 3. "AI primarily improves the variety and depth of ideas, rather than their fluency or usefulness" (abstract). No effect sizes available. - **Relevance:** (a) — | (b) implies facilitation of team elaboration/convergence determines AI benefit | (c) generative AI raises idea diversity in teams; novelty gains conditional on team information elaboration; no gain in fluency or usefulness - **Caveats:** abstract only; no statistics, effect sizes, AI tool or task details; field experiments described only in one sentence; possible later full publication not checked. - **Screening check:** meets criteria on the abstract (empirical, teams of humans with generative AI, n reported), but method details are unverified — flag as abstract-level evidence. #### R685 — Unequal GenAI access in two-person teams - **Citation:** Han, J., & Ren, R. (2025). Why unequal AI access enhances team productivity: the mediating role of interaction processes and cognitive diversity. Frontiers in Psychology, 16, 1636906. https://doi.org/10.3389/fpsyg.2025.1636906 - **Type:** journal — peer-reviewed: yes - **Full text read:** yes (https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2025.1636906/pdf; pages = PDF pages) - **Method:** randomised controlled lab experiment; each team first did a control writing task without GenAI, then was randomised to unequal access (one member) or full access (both) for a second task; interaction coded (Cohen's kappa 0.77); OLS and PROCESS mediation (§3, p. 5-8) - **n:** 120 university students in 60 two-person teams (p. 5-6) - **Setting:** face-to-face press-release writing tasks (≤45 min) in China; LLM Kimi 3.0 (Moonshot) (p. 6) - **Findings:** 1. Task quality was higher with unequal access (M = 5.654) than with no access (M = 4.981; t = -5.103, p < 0.01) and full access (M = 5.250; t = 2.881, p < 0.01) (§4.1, p. 8). 2. Completion time was shorter with unequal access (M = 10.013 min) than no access (M = 30.483; t = 13.556, p < 0.01) and full access (M = 15.150; t = -3.312, p < 0.01) (§4.1, p. 8). 3. The quality advantage over full access ran through cognitive diversity (indirect effect b = -0.283, 95% CI [-0.513, -0.134]); unequal access increased concentrated questioning, which shortened task time (p. 9-10). - **Relevance:** (a) — | (b) — | (c) how GenAI access configuration changes team interaction, cognitive diversity and output quality - **Caveats:** student sample; dyads; writing task (not ideation/decision session); no-access condition is a within-team control phase preceding the treatment (order confound). - **Screening check:** meets criteria — randomised experiment, two humans interacting as a team with an LLM, method and n stated, English. #### GS15 — DALL-E vs Miro in group brainstorming - **Citation:** Gerardini, K., Sarcinella, E. D., Borghesi, F., Pozzi, A., Gaggioli, A., & Chirico, A. (2025). Supporting Brainstorming with a Collaborative Platform or a Generative Artificial Intelligence Tool: An Exploratory Study. Cyberpsychology, Behavior, and Social Networking, 28(7), 524–533. https://doi.org/10.1089/cyber.2024.0518 (PMID 40495699) - **Type:** journal — peer-reviewed: yes - **Full text read:** abstract only (Crossref/OpenAlex abstract; publisher PDF closed access, Europe PMC not OA) - **Method:** 2 × 2 experimental design (online vs offline; Miro vs DALL-E); measures: PANAS affect, Aesthetic Emotions Questionnaire, creative self-efficacy, Technology Readiness Survey, User Experience Questionnaire, Flow State Scale–Short, creativity (fluidity, elaboration, flexibility, originality) rated by two independent raters (abstract). Whether tool factor was within- or between-groups not verified. - **n:** 16 groups of 5 members (= 80 participants, derived from abstract) (abstract). - **Setting:** Group brainstorming supported either by Miro (conventional whiteboard) or DALL-E (image-generating GenAI), online or offline; participant population not verified. - **Findings:** 1. DALL-E produced more positive affect, richer aesthetic experiences, and higher attractiveness and novelty (UX) than Miro, particularly online (abstract). 2. Significant interaction effects for "efficiency" (UX) and "union action-consciousness" (flow) (abstract; statistics not verified). 3. Participants felt more creative and preferred DALL-E; online sessions with DALL-E led to greater idea elaboration (abstract). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: GenAI image tool as brainstorming support medium | (c) AI and group ideation/decisions: effects on group creativity (elaboration) and experience - **Caveats:** Abstract only; exploratory; effect sizes not verified; same research team published related studies — MetroXRAINE 2024 pilot (https://doi.org/10.1109/metroxraine62247.2024.10796552, ChatGPT/DALL-E/Miro), CCIS 2025 preliminary study on aesthetic emotions (https://doi.org/10.1007/978-3-032-05778-5_6), and LNICST 2026 feasibility study (https://doi.org/10.1007/978-3-032-26235-6_10, = OA1567) — samples may overlap; avoid double counting. - **Screening check:** Meets criteria per abstract — empirical experiment with stated design and n (16 groups × 5), ≥2 humans brainstorming as a group with generative AI (DALL-E), English. #### OA1419 — AI image stimuli in group ideation (EPDE 2025) - **Citation:** Gong, Z., Wang, M., Soomro, S. A., Paananen, S., Nurmela, P., Häkkilä, J., & Georgiev, G. V. (2025). Influence of image stimuli on design creativity: Exploration of generative AI in group ideation. Proceedings of the 27th International Conference on Engineering and Product Design Education (E&PDE 2025), University of Malta, pp. 19–24. https://doi.org/10.35199/epde.2025.4 - **Type:** conference — peer-reviewed: unclear (Design Society conference proceedings; review process not stated in the paper) - **Full text read:** yes (https://www.designsociety.org/download-publication/48545/influence_of_image_stimuli_on_design_creativity_exploration_of_generative_ai_in_group_ideation) - **Method:** in-class exercise (field, not controlled; no comparison condition): collaborative sketching in groups of 3 with one pre-generated AI image stimulus per participant; external raters judged novelty/usefulness (1–5) and source of contribution of each final idea; thematic analysis of written reflections (§3, p. 2–3 of PDF). - **n:** ~30 participants took part, 18 analysed (incomplete exercises excluded); groups of three; 6 external evaluators (each idea rated by 2) plus the instructor; 18 final ideas; 15 written reflections (§3.1, §3.4.1, §4.3; p. 2–4 of PDF). - **Setting:** master's/PhD students (mean age 25.06, SD 2.64; 14 male, 2 female, 2 undisclosed) in a creative design course (Univ. of Lapland/Oulu, Finland); task: design an inclusive playground for wheelchair users; images generated beforehand with "AI image generators, such as Dall-E and Midjourney" using the task description as prompt; participants did not interact with the AI themselves (§3.1–3.3, §5.5; p. 2, 5 of PDF). - **Findings:** 1. Source of the novelty of the 18 final ideas: 6 from group members, 1 from the AI image stimulus, 2 from both, 9 from participants' own initial ideas; for usefulness: 14 group members, 2 both, 2 neither; no difference in novelty/usefulness scores across contribution sources (§4.2, Table 1; p. 4 of PDF). Authors summarise that "only 6 out of 18 final ideas benefitted from the generative AI image stimuli" (§5.1, p. 4 of PDF). 2. Of 15 reflections, 8 mentioned the visual stimuli, "with the majority expressing the visual stimuli as an inspiration for generating ideas is meaningless"; only 1 said the stimulus helped; 11 credited their group members with developing/improving their ideas; 10 said the group method made the exercise fun and instructive (§4.3, p. 4 of PDF). 3. Inter-rater agreement across three evaluations: Kendall's W = .532, p = .014 (§4.1, p. 3 of PDF). - **Relevance:** (a) — | (b) weak: static AI-generated stimuli inserted into a facilitated collaborative-sketching session | (c) AI image stimuli contributed little to group ideation vs. peer input - **Caveats:** small n (18), no control condition, student sample, uncontrolled classroom setting, AI stimuli pre-generated (no human–AI interaction). Internal inconsistency: Table 1 lists 6 novelty ideas from group members, but §5.2 (p. 4–5 of PDF) says "seven participants were inspired by group members" for novelty and "16 out of 18" used group input for usefulness (Table 1: 14 + 2 = 16). The text also refers to "18 patients" in §4.1 (typo). - **Screening check:** meets criteria marginally — empirical, n reported, 3-person human groups with generative-AI (image) involvement, though AI was only a pre-generated stimulus, not an interactive participant. #### OA1438 — Text-to-image AI as catalyst for semantic convergence (DIS 2024) - **Citation:** Lin, P.-Y., Andersen, K., Schmidt, R., Schoenmakers, S., Hofmeyer, H., Pauwels, P., & IJsselsteijn, W. (2024). Text-to-image AI as a catalyst for semantic convergence in creative collaborations. In Proceedings of the 2024 ACM Designing Interactive Systems Conference (DIS '24), pp. 2753–2767. ACM. https://doi.org/10.1145/3643834.3661543 - **Type:** conference — peer-reviewed: yes (ACM DIS full paper; repository copy is the Version of Record) - **Full text read:** yes (https://pure.tue.nl/ws/portalfiles/portal/334025468/3643834.3661543.pdf) - **Method:** qualitative research-through-design workshops: 30-min facilitated sessions (5-min pre-questionnaire + 25-min interactive think-aloud session around a shared large screen), screen and audio recording, semi-structured interview, post-workshop questionnaire; thematic analysis (Braun & Clarke) plus interaction analysis of transcripts; sessions run "until repeating patterns were observed" across two semesters (§2.2–2.4, pp. 2756). - **n:** 27 participants in 14 sessions: 5 team sessions (18 students, teams of 3–4) and 9 individual sessions (§2.2, p. 2756). - **Setting:** undergraduate/master industrial design students in a semester-long "wearables" project course (TU Eindhoven), in the ideation phase; Midjourney v3/v4 via Discord, Sept 2022–March 2023; a facilitator operated the interface and typed prompts while participants sat around the screen (§2.1–2.3, pp. 2755–2756). - **Findings:** 1. Only 11 of 27 participants had prior text-to-image AI experience; all 27 found the workshop helpful, but only 2 said the AI generated an image representing their project, while the other 25 said it instead gave them more options and ideas (§3, pp. 2756–2757). 2. Idea alignment between participants and the AI through iterative prompt rewording (verbal articulation) occurred in 10 of 14 sessions (team and individual) (§3.2, p. 2757). Two participants from different teams wrote that "Midjourney helps them to see what others (teammates) see" and two others that "AI helps to align the group (team) vision" (§3, p. 2757). 3. In team sessions the generated images became the focus of discussion: members negotiated interpretations of ambiguous images toward agreement, pushing discussion of actual design ideas into a "secondary discussion space" attended to later; all participants reported the exploration did not define their project's final vision, and the executability of AI-suggested ideas was "generally considered low" (§3.2–3.3, pp. 2758–2759; §4.4). - **Relevance:** (a) — (a facilitator ran the AI, but facilitator practice is not studied; post-questionnaire Q6 found no participant reported the facilitator as inspiring or guiding) | (b) text-to-image AI as a shared artefact in facilitated design-ideation workshops | (c) AI images supported semantic convergence/alignment in teams, less so final idea generation - **Caveats:** qualitative, no comparison condition, student sample, only 5 team sessions (18 people) are group settings; older Midjourney versions (v3/v4); counts come from questionnaires, not measured outcomes. - **Screening check:** meets criteria — empirical, n reported, teams of 3–4 humans interacting with generative AI in a facilitated session. #### C542 — AtoMix: GenAI-augmented remote 6-3-5 visual brainwriting - **Citation:** Kim, H. J., Khan, A. A., Huh, S., & Bianchi, A. (2026). AtoMix: Fostering Structured Visual Ideation for Remote Groups through Atomic Composition and Cross-Pollination. In Proceedings of the 2026 Conference on Creativity and Cognition (C&C '26), pp. 1092–1107. ACM. https://doi.org/10.1145/3803784.3807530 - **Type:** conference — peer-reviewed: yes - **Full text read:** yes (author PDF, https://make.kaist.ac.kr/files/2026/KimHeeji_AtoMix_CC26.pdf; CC BY 4.0; page numbers = PDF pages 1–16) - **Method:** (1) Formative study: three 2-hour 6-3-5 visual brainwriting sessions in Figma (6 rounds × 10 min) with free use of ChatGPT/Gemini/Copilot, followed by 40-min group interviews, thematic analysis (§3.1, p. 4). (2) Between-subjects comparative study: four remote groups of six (2 AtoMix, 2 baseline), 30-min 6-3-5 session (5-min rounds) + 30-min group interview; action logs, LLM-as-judge ratings of outputs (claude-sonnet-4-5, 20% re-coded by a researcher, 92.7%/89.2% agreement), Mann–Whitney tests, thematic analysis (§5, pp. 5–8). - **n:** Formative N = 18 (3 groups of 6); comparative study N = 24 (4 groups of 6); 108 canvases per group; 72 vertical and 144 horizontal evaluation cases (§3.1, p. 4; §5.1–5.3, pp. 6–7). - **Setting:** University students/researchers with design or HCI experience (comparative study mean age 25.4), remote via video call; task "Design experiences for a learning environment in virtual reality"; image generation with Gemini 2.5 Flash (Image) in both conditions; baseline = separate generator with whole images copied onto canvas (§5.1, p. 7; §4.2, p. 5). - **Findings:** 1. Formative study: even with GenAI embedded in 6-3-5, participants fell into prompt–regenerate loops treating outputs as finished images, struggled to express multi-component ideas and build on others' images, "leading to redundancy and limited divergence" (abstract; §3.2, pp. 2, 4–5). 2. AtoMix vs. baseline: images per canvas M = 9.5 vs. 4.42 (p < .001); horizontal idea differentiation (LLM-judged) visual distinctiveness 4.61 vs. 3.67 (p = .0007, d = 0.696) and conceptual diversity 4.35 vs. 3.53 (p = .0037, d = 0.608); vertical continuity 3.72 vs. 3.44 and idea progression 3.56 vs. 3.19 not significant (§6.1, pp. 8–9; Table 1, p. 9). 3. Empty canvases 37/216 (17.1%) vs. 62/216 (28.7%) (p < .01); AtoMix users submitted more (M = 15.25 vs. 8.58) and shorter prompts (3.26 vs. 10.93 words), with 58 cross-pollination actions; baseline participants more often described the task as "generating ideas individually rather than developing ideas together" (§6.1.4, p. 8; §6.2, pp. 9–10; §6.3.2, pp. 11–12). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: generative-image support embedded in a structured group ideation method (6-3-5 brainwriting) for remote groups | (c) AI and group ideation/decisions: interface design (fragment-level vs. whole-image generation) changed idea diversity, build-on and canvas completion in group brainwriting - **Caveats:** Small n (2 groups per condition, groups as units not analysed); student/researcher sample; outcome quality rated by an LLM judge with single-coder spot check; system integration confounded with compositional features (acknowledged, §7.1, p. 12). - **Screening check:** Meets criteria — empirical comparative study with groups of 6 humans doing structured group ideation with generative AI, method and n reported, English. #### C496 — Human agency vs. AI autonomy in human-AI idea selection (IdeaMap) - **Citation:** Young-Ng, M., Zhu, Q., Liao, J., & Wang, H.-C. (2025). Balancing Human Agency and AI Autonomy in Human-AI Idea Selection. In Extended Abstracts of the CHI Conference on Human Factors in Computing Systems (CHI EA '25), pp. 1–6. ACM. https://doi.org/10.1145/3706599.3719880 - **Type:** conference (extended abstract / late-breaking work) — peer-reviewed: yes (lighter review, EA) - **Full text read:** abstract only (ACM PDF blocked by Cloudflare; no arXiv or author copy found; OpenAlex abstract field contains only the Figure 1 caption, so abstract content below is taken from the ACM landing-page abstract as surfaced by web search — treat as not fully verified) - **Method:** Participatory design workshop using IdeaMap as a technology probe to elicit feedback on desired AI features for convergent thinking (idea selection) in single- and multi-party group brainstorming. - **n:** 12 participants (workshop); group composition/number of sessions not verified. - **Setting:** IdeaMap: human-AI group brainstorming system with a chatroom, sidebar of usernames, and an ideation-map visualisation of posted ideas as a shared workspace; LLM gpt-3.5-turbo. - **Findings:** 1. Workshop uncovered user perceptions on human agency and AI autonomy in human-AI teaming for convergent thinking (idea selection) (abstract; specific findings not verified). 2. Framed around Human-Centered AI: balancing automation with user contribution oversight is key to preserving ownership in human-AI co-creation (abstract). 3. No quantitative results available from the abstract. - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: design requirements for AI support of the idea-selection/convergence phase in group brainstorming | (c) AI and group ideation/decisions: perceptions of how much autonomy AI should have in group idea selection - **Caveats:** Abstract only, 6-page extended abstract, small n, design-elicitation (technology probe) rather than outcome evaluation. - **Screening check:** Probably meets criteria (empirical workshop, n = 12, LLM-based group brainstorming probe incl. multi-party use, English) — whether participants actually used IdeaMap as groups is not verified; confirm from full text. #### OA1341 — Rapid AIdeation (LLM in a rapid-ideation workshop) - **Citation:** Lim, G., & Perrault, S. T. (2024). Rapid AIdeation: Generating Ideas With the Self and in Collaboration With Large Language Models. arXiv preprint arXiv:2403.12928. https://doi.org/10.48550/arxiv.2403.12928 - **Type:** preprint (short workshop-style paper, 6 pp. + appendix) — peer-reviewed: no (preprint) - **Full text read:** yes (https://arxiv.org/pdf/2403.12928, v1) - **Method:** Exploratory online design workshop (40 min, video conferencing) with fixed order: individual "Self Ideation" (Crazy 8's) → individual "Co-GPT Ideation" with ChatGPT → group Affinity Diagramming (KJ method) → group Ranking of top three ideas; exploratory quantitative content analysis of post-its, descriptive statistics of rankings, qualitative content analysis of prompts (§3.1, §3.3, p. 2–3). IRB approved (§3, p. 2). - **n:** 21 university students in random groups of three: 7 groups (4 on "Misinformation", 3 on "Hallucination"); 108 Self-Ideation post-its, 119 Co-GPT post-its; 59 prompts from 17 voluntarily submitted ChatGPT conversations (§3.2, §4, p. 3). Demographics from 15 optional survey respondents (9 women, 6 men) (§3.2, p. 3). - **Setting:** University in Singapore (SUTD affiliation); online workshop; task: ideas to use ChatGPT against online misinformation / to address ChatGPT hallucination; ChatGPT v3.5 (§3.1, p. 2; §6, p. 5). Organizer acted mainly as timekeeper (§3.1, p. 2). - **Findings:** 1. 23.6% more unique ideas with Co-GPT Ideation than Self Ideation (47 vs 38; Misinformation 26 vs 21, Hallucination 21 vs 17) (§4.1, p. 3–4). 2. Groups' top-three representative ideas: 12 from Co-GPT vs 9 from Self Ideation; Co-GPT post-its were longer (M = 146.8 chars, SD = 86.8 vs M = 97.7, SD = 50.7); but more Self-Ideation post-its ranked first (4 vs 3), while Co-GPT post-its ranked more often second (4 vs 3) and third (5 vs 2) — authors: quality question "not a straightforward one" (§4.2, Table 1, p. 4). 3. Prompts were short and straightforward (avg 3.59 prompts, SD = 2.37; 86.9 characters, SD = 73.9); no participant used advanced prompting such as examples; two participants used an antagonistic "anti-collaboration" style (threatening/interrogating ChatGPT) (§4.3, p. 4; §5.3–5.4, p. 5). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: LLM used as an ideation input stage inside a facilitated rapid-ideation workshop protocol | (c) AI and group ideation/decisions: more unique ideas with LLM; group selection of ideas did not clearly favour LLM ideas at rank 1. - **Caveats:** Preprint; small student sample; fixed order (Self before Co-GPT), no counterbalancing (§6, p. 5); "unique"/"high quality" loosely defined, quality = group's own selection; ChatGPT use was individual, only clustering/ranking was done in groups; GPT-3.5. - **Screening check:** Meets criteria (empirical; groups of three humans ideate with LLM-generated input and jointly select ideas; method and n reported; English). Borderline because LLM interaction itself was individual, not group-level. #### OA1047 — Teachers ideating in groups with a Design-Thinking LLM multi-agent system - **Citation:** Zampolini, L., Mangione, G. R. J., & Gentile, M. (2026). Supporting Design Thinking through Ideation LLM-Multi-Agent Systems: An Investigation of Perceived Usefulness in Non-Standard Educational Contexts. Proceedings of the 18th International Conference on Computer Supported Education (CSEDU 2026), Vol. 1, pp. 346–357. SciTePress. https://doi.org/10.5220/0014816500004021 - **Type:** conference — peer-reviewed: unclear (CSEDU proceedings; review not described in text) - **Full text read:** yes (https://doi.org/10.5220/0014816500004021; SciTePress PDF) - **Method:** Single 90-minute hands-on workshop (training + group brainstorming in multi-user chat rooms with an LLM multi-agent system), followed by individual self-report questionnaire (four 6-item constructs, 5-point Likert); descriptive statistics, Pearson correlations, path analysis (WLSMV), bootstrap mediation (5,000 resamples). No comparison condition (§3.1–3.5, p. 351–353). - **n:** 24 primary and lower-secondary teachers in 4 groups/virtual rooms (6, 5, 7, 6 humans), 2 human facilitators per room; 20 completed the questionnaire and were analysed (§3.1, Table 1, p. 351–352). 7 candidate challenges; brainstorming sessions of 5–10 minutes in multiple cycles (§3.2, p. 351). - **Setting:** Workshop "Ideating with AI" at Didacta Italia (Trentino edition), October 2025, challenges from small and rural schools; LangChain-based LLM-MAS with 5 proactive agents (an expert designer "that acts as a facilitator", Conceptualizer, Gardener, Jester, Challenger) plus a synthesiser agent summarising each session; underlying LLM described only as "a general-purpose language model" (model not verified) (§3.3, p. 352; §5.3, p. 355). Participants interacted "exclusively through the LLM-MAS chat, with only minor external oral discussion"; researchers facilitated and used the synthesiser agent to summarise (§3.2, p. 351). - **Findings:** 1. Correlations: AI knowledge–AI trust r = .71; system perceived usefulness (SYS_PU) with AI trust r = .54 and with DT knowledge r = .47; AI knowledge–SYS_PU r = .37 (§4, p. 353). 2. Path model (χ²(1) = 0.82, p = .364; CFI = 1.000): AI knowledge → trust β = .70 (p = .005); DT knowledge → SYS_PU β = .41 (p = .013); trust → SYS_PU β = .41 (p = .056, marginal); AI knowledge → SYS_PU direct β ≈ .00 (p = .985); model explained ~39% of SYS_PU variance; indirect effect via trust not significant (β = .31, p = .387); DT-knowledge effect "unstable" under bootstrap (§4, p. 353–354). 3. Authors' reflection: agent cooldown/reactivity were not calibrated and "overly reactive agents may limit human–human interaction and overshadow participants' contributions"; questionnaire not linked to interaction traces (§5.3, p. 355). No mean usefulness score reported in text (not verified). - **Relevance:** (a) facilitator use of AI: human researchers facilitated sessions and used an AI synthesiser to summarise rounds | (b) AI planning/running/supporting group sessions: LLM agents including a facilitator-role agent and a synthesiser inside group ideation rooms | (c) AI and group ideation/decisions: only perceived usefulness measured; no ideation output or decision-quality data. - **Caveats:** Very small sample (20 analysed) for path analysis; self-report only; no control condition; conference workshop volunteers; LLM unspecified; authors developed the system. - **Screening check:** Meets criteria (empirical; groups of 5–7 teachers interact with LLM agents in shared chat; method and n reported; English). #### OA1099 — Human-AI role allocation in Design Thinking (pairs + ChatGPT) - **Citation:** Rodenas, I., & Rupprecht, P. (2026). Human-AI Role Allocation in Design Thinking: Contributions to Initiation, Elaboration, and Decision-Making. In Human Interaction and Emerging Technologies (IHIET 2026), AHFE International, Vol. 236, pp. 259–268. https://doi.org/10.54941/ahfe1008101 - **Type:** conference — peer-reviewed: unclear (AHFE proceedings; "Received May 26, 2026; Revised July 1, 2026; Accepted July 17, 2026" suggests review, p. 259) - **Full text read:** yes (https://openaccess-api.cms-conferences.org/articles/download/979-8-950676-12-3_25) - **Method:** Exploratory randomized two-condition experiment; pairs completed a 60-minute Design Thinking task (Empathise & Define, Ideate & Decide, Prototype, Test) with ChatGPT under two role-allocation sequences: Team A AI–Human–AI (AI initiates and produces final version/preselection) vs Team B Human–AI–Human (humans initiate, AI expands/structures, humans decide). Measures: SUS, NASA-RTLX (physical demand excluded), open-ended responses, chat transcripts. Descriptive statistics only, no inferential tests (p. 262–263). - **n:** Abstract: 32 working professionals; Method section: 33 participants, questionnaire data 16 (Team A) + 17 (Team B) (p. 259 vs p. 261). Worked in pairs; number of pairs not stated (not verified). - **Setting:** Working professionals in the "Business Planning" course at FHWien der WKW, Vienna; task: product concept for a fictitious Vienna company's hydroponic planting system; ChatGPT version 5.2 via separate ChatGPT projects per condition; Google Search allowed in human phases (p. 262). - **Findings:** 1. Mean SUS 79.84 (AI–Human–AI) vs 70.59 (Human–AI–Human) (p. 263, Fig. 1). 2. NASA-RTLX subscales Team A vs Team B: mental 26.90 vs 22.00, performance 32.41 vs 27.30, frustration 34.48 vs 24.70, temporal 28.28 vs 18, effort 28.28 vs 24.70; authors state "higher scores consistently indicating a lower perceived workload" and interpret AI–Human–AI as lower workload (p. 264). Note: this scoring direction is unusual for RTLX and is reported as stated. 3. Qualitative: in AI–Human–AI, AI acted as a "structuring and partly steering process actor", with concerns about "AI dominance and loss of control"; in Human–AI–Human, concerns were output length, limited originality, overly affirmative responses and need for revision — authors frame a trade-off between cognitive efficiency and human control/ownership; reuse intention 12 (Team A) vs 13 (Team B) respondents (p. 263, 265–266). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: tests who (AI vs humans) initiates and closes each Design Thinking phase — a process-design/facilitation choice | (c) AI and group ideation/decisions: AI-led sequencing felt easier but raised ownership/control concerns; no outcome-quality measure. - **Caveats:** Descriptive only, small sample, n inconsistency (32 vs 33), single task, non-standard RTLX scoring/direction, no quality rating of concepts; pairs (dyads) rather than larger groups. - **Screening check:** Meets criteria (empirical; pairs of professionals collaborate with ChatGPT; method and n reported; English). Borderline on group size (dyads) but ≥2 humans. #### R0 — When Teams Embrace AI (co-prompting in pairs) - **Citation:** Han, Y., Qiu, Z., Cheng, J. S. F., & LC, R. (2024). When Teams Embrace AI: Human Collaboration Strategies in Generative Prompting in a Creative Design Task. Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (CHI '24), ACM, pp. 1-14. https://doi.org/10.1145/3613904.3642133 - **Type:** conference — peer-reviewed: yes - **Full text read:** yes (arXiv v1, https://arxiv.org/pdf/2509.21731; page numbers below are pages of the arXiv PDF) - **Method:** qualitative workshop study; screen-recorded co-prompting sessions, semi-structured interviews, thematic coding (Sec. 3, p. 3-5) - **n:** 18 participants (8 male, 10 female) randomly assigned to 9 pairs (p. 3) - **Setting:** undergraduate-level students with design/performance backgrounds, little GenAI experience; pairs created stage designs from a poem using Midjourney (text-to-image) and optionally ChatGPT (GPT-3.5) in an online meeting with screen sharing (p. 3-4) - **Findings:** 1. Two collaboration workflows emerged: plot-based (Groups 2, 3, 8, 9) and concept-based (Groups 1, 4-7); ChatGPT "seemed to be used more" in plot-based groups (§4.1, p. 5). 2. GenAI outputs became "a basis for negotiation" in divergent views, easing consensus; e.g., one pair used ChatGPT to turn their conflicting ideas into stories and then reached consensus through discussion of the outputs (§4.3.3, p. 8-9, Fig. 11). 3. The majority perceived GenAI as a tool rather than a collaborative partner, citing its inability to engage in discussion or team decision-making; participants took curator/director roles and delegated laborious tasks to GenAI (§4.4.1, p. 9). - **Relevance:** (a) — | (b) GenAI as a shared artefact supporting small-group design sessions | (c) GenAI outputs helped pairs converge/build consensus; qualitative only - **Caveats:** small student sample, pairs only (n=2 per group), qualitative/no comparison condition; arXiv version posted 2025, published CHI 2024. - **Screening check:** meets criteria — empirical qualitative study, 9 two-person groups interacting with GenAI, method and n stated, English. #### R964 — LLMs in co-design team collaboration and creativity - **Citation:** Zhang, D., Shen, J., Luo, S., Zhang, X., & Liu, W. (2026). Reshaping team collaboration and creativity in co-design: the role of large language models. CoDesign, 1-36. https://doi.org/10.1080/15710882.2026.2641141 - **Type:** journal — peer-reviewed: yes - **Full text read:** abstract only (not open access; T&F blocked) - **Method:** mixed-methods study combining protocol analysis, self-report scales and expert evaluation, analysed with a co-evolution-model framework (Abstract) - **n:** N = 40 (Abstract); number of teams and team size not verified - **Setting:** novice designer teams doing LLM-supported co-design (LLM model not verified) - **Findings:** 1. LLMs accelerated transitions between problem and solution spaces, increasing the pace of collaboration, and were associated with fewer segments per space unit (Abstract; numbers not verified). 2. LLMs supported information processing and early-stage ideation in the problem space, but in the solution space LLM use was associated with lower engagement in conceptual integration, functional reasoning and decision-making (Abstract). 3. LLMs enhanced outcome novelty and self-perceived creativity but showed limited improvement in outcome usefulness and in team-level creativity support (expressiveness, collaboration, immersion) (Abstract). - **Relevance:** (a) — | (b) — | (c) LLM support in team co-design raises novelty/pace but reduces convergent integration and decision-making engagement - **Caveats:** abstract only, no effect sizes; novice (likely student) sample; overlapping author team with R1506 (Zhang, Luo, Shen, Zhang, Liu) — possible related data, not verified - **Screening check:** meets criteria — empirical mixed-methods study, human teams collaborating with an LLM, n reported (N = 40), English. #### R1506 — LLMs and collective cognition in collaborative new-product development - **Citation:** Zhang, D., Luo, S., Liu, Y., Zhang, X., Hu, Y., Shen, J., Yi, P., Zhao, K., & Liu, W. (2026). Large language model tools as catalysts for collective cognition in collaborative new-product development: a quasi-experimental study. Humanities and Social Sciences Communications, 13, 382. https://doi.org/10.1057/s41599-026-06738-7 - **Type:** journal — peer-reviewed: yes - **Full text read:** yes (https://www.nature.com/articles/s41599-026-06738-7.pdf) - **Method:** quasi-experiment (LLM-assisted vs non-LLM teams) with think-aloud protocol analysis coded by the 2I2A model (team-level analyses), plus semi-structured interviews analysed with grounded theory (pp. 5, 7-8) - **n:** 44 participants in 22 two-person teams, 11 teams per condition; interviews with the 22 participants of the LLM condition (pp. 5, 8) - **Setting:** mostly undergraduate/junior master's students (39 undergraduates = 88.64%, Table 1, p. 5) designing an innovative product for children in 90-min sessions; LLM condition used ChatGPT-4.0 or Kimi, control had no digital devices (p. 5) - **Findings:** 1. LLM teams collaborated more in the Identification Space (M = 18.64 vs 10.00, t(13.297) = -9.219, p < 0.001, d = -3.93) and Interaction Space (M = 28.18 vs 18.64, t(20) = -8.224, p < 0.001, d = -3.51); no differences in Analysis (p = 0.467) or Accommodation (p = 0.937) Spaces (Table 4, p. 9). 2. Non-LLM teams showed more Evaluation (M = 35.36 vs 26.45, F(1,20) = 18.217, p < 0.001, partial eta2 = 0.477) and Negotiation and Interpretation (M = 44.36 vs 34.73, F(1,20) = 10.721, p = 0.004); no difference in Creative Output (11.18 vs 12.73, p = 0.186) (p. 9). 3. Interviews point to reduced collaborative naturalness and a tendency toward over-reliance on LLM suggestions (Abstract; grounded-theory results pp. 11-14). - **Relevance:** (a) — | (b) — | (c) LLM access expands early-stage idea/information generation in teams but reduces evaluation and negotiation, with no gain in creative output - **Caveats:** dyads only; student sample; quasi-experimental; two different LLMs pooled; overlapping authors with R964 (possible related data, not verified) - **Screening check:** meets criteria — empirical quasi-experiment with 2-person human teams using LLMs in a joint design task, method and n reported, English. #### R672 — AI-Mediated Group Ideation (architecture) - **Citation:** Gan, A. W. J., Dang, Q., Western, B., & García del Castillo y López, J. L. (2023). AI-Mediated Group Ideation: Framework for collaboration in design. eCAADe 2023 Proceedings, Vol. 2, pp. 389-398. https://doi.org/10.52842/conf.ecaade.2023.2.389 - **Type:** conference — peer-reviewed: unclear (eCAADe papers are reviewed; not verified for this paper) - **Full text read:** yes (http://papers.cumincad.org/data/works/att/ecaade2023_444.pdf; pages as printed, 389-398) - **Method:** design framework and prototype plus a single case-study session with post-session participant reflections (informal, qualitative) (p. 395-396) - **n:** 1 group of 4 graduate students (architecture and unrelated fields), one 30-minute virtual session (p. 395) - **Setting:** web app "AI-mediator" for architectural group ideation: real-time speech-to-text (AssemblyAI), periodic summaries, text-to-image generation (OpenAI/DALL-E) with in-painting, a speaking-time "conversational clock", posture alerts (p. 393, 395) - **Findings:** 1. Participants found the generated image prompts useful for ideating; generated images that felt "somewhat conventional" prompted the group to push the design further (p. 395-396). 2. Images were presented without hierarchy, causing memorability problems (p. 396). 3. A 3-minute summary cadence was "too frequent", resulting in distraction (p. 396). - **Relevance:** (a) — | (b) prototype AI mediator for facilitated group ideation (summaries, participation balance) | (c) generated images stimulated/redirected group ideation (anecdotal) - **Caveats:** single group of 4, no systematic data collection or measures reported; primarily a design/framework paper. - **Screening check:** marginally meets criteria — a group session with generative AI, method (case study) and n stated; evidence is anecdotal. Could be argued EX-NOTEMP (prototype with minimal evaluation). #### C25 — Cognitive Bridge: AI-generated boundary objects for cross-functional collaboration - **Citation:** Gunasekaran, T. S., Lim, S., Gupta, K. K., Bai, H., Pai, Y. S., & Billinghurst, M. (2026). Cognitive Bridge: AI-Generated Boundary Objects for Cross-Functional Collaboration. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26), pp. 1–35. https://doi.org/10.1145/3772318.3791399 - **Type:** conference (full paper) — peer-reviewed: yes - **Full text read:** abstract only (ACM DL blocked; no preprint found) - **Method:** Formative study with professionals, system design, and a controlled study comparing Cognitive Bridge with baseline tools (Abstract). Study design details (within/between, statistics) not verified. - **n:** Formative study: 7 professionals; controlled study: 16 designer–developer dyads (Abstract). - **Setting:** Designer–developer pairs collaborating on a whiteboard; AI system monitors multimodal cues (facial expressions, speech, workspace activity) to detect emerging misunderstandings and generates adaptive boundary objects (visual diagrams, wireframes, flowcharts) in real time; underlying model not stated in abstract (not verified). - **Findings:** 1. Compared with baseline tools, Cognitive Bridge reduced communication conflicts by 47% and increased implementable solutions by 34% (Abstract; statistical tests not verified). 2. Solution–exploration trade-off: AI accelerated alignment but risked premature convergence that constrained creative exploration (Abstract). - **Relevance:** (a) — | (b) real-time AI support that detects misunderstanding and produces shared artifacts during collaborative sessions (a facilitation-like "translation" function) | (c) faster alignment and more implementable solutions, but premature convergence reducing exploration - **Caveats:** Abstract only; effect figures unverified against full text; lab setting, small n; generative-AI component inferred from description ("AI-generated" diagrams/wireframes) but model not verified. - **Screening check:** Likely meets criteria (empirical, dyads collaborating with AI-generated artifacts, method and n reported); LLM/generative-AI involvement not explicitly confirmed in abstract — not verified. #### C59 — XSynth: GenAI-built shared mental models for XR design teams - **Citation:** Li, Y., Li, S., He, Z., & Wang, D. (2026). XSynth: GenAI-Empowered Shared Mental Model Building for Conceptual Design Collaboration in Extended Reality. Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26), pp. 1–24. https://doi.org/10.1145/3772318.3790350 - **Type:** conference — peer-reviewed: yes - **Full text read:** abstract only. ACM DL is behind a Cloudflare challenge, and no arXiv or author copy was found. Abstract from OpenAlex, plus an ACM DL search snippet stating that the GenAI agent uses ByteDance's Doubao LLM. - **Method:** Formative identification of cognitive challenges, then a within-subject experiment (XSynth vs. a baseline; the baseline is not specified in the abstract) with mixed methods (abstract). - **n:** 10 design teams, N = 30 (abstract). - **Setting:** Automotive conceptual design collaboration in XR. Each designer builds an individual mental model with a GenAI agent (Doubao LLM, per the ACM DL snippet, not verified in the full text). The agent merges the individual models into a unified knowledge graph (a shared mental model), grounded in Concept–Knowledge theory (abstract). - **Findings:** 1. XSynth "significantly reduced workload, enhanced creativity support, strengthened perceived SMMs, and improved design performance" (abstract). Exact statistics were not verified. - **Relevance:** (a) — | (b) GenAI synthesis of individual views into a shared model to align a design team (a convergence/alignment facilitation function) | (c) better perceived creativity support and design performance (numbers not verified) - **Caveats:** Abstract only, so no effect sizes, baseline or model details were verified. Small n (10 teams). - **Screening check:** Likely meets criteria (empirical, teams of about 3 humans collaborating with GenAI, n reported, English). The group-interaction details could not be verified from the full text. #### C34 — Shared gaze on AI node maps: mitigating "empathy fog" in AI-mediated dyadic discussion - **Citation:** Yamashita, J., Saito, H., Sakurada, K., Takashita, S., Muramoto, G., & Inami, M. (2026). Shared Gaze on AI Node Maps: Mitigating Empathy Fog in AI-Mediated Dyadic Discussion. In Proceedings of the Augmented Humans International Conference 2026 (AHs '26), pp. 202–212. https://doi.org/10.1145/3795011.3795058 - **Type:** conference (full paper) — peer-reviewed: yes (not verified beyond venue) - **Full text read:** abstract only (ACM DL blocked; no preprint found) - **Method:** Exploratory laboratory study, three within-subject conditions: face-to-face, AI node map only, AI node map + Shared Gaze Visualization (SGV) (Abstract). - **n:** 16 participants in 8 dyads (Abstract). - **Setting:** Dyads solving Hidden Profile tasks; LLM-based AI-mediated communication system summarises, rephrases and structures dialogue into a shared node map (specific model not stated in abstract — not verified). - **Findings:** 1. Compared with face-to-face, the map-only condition was associated with a shift of the locus of interaction toward the tool and an absence of creative hypotheses ("Empathy Fog") (Abstract). 2. Adding shared gaze increased perceived surveillance without inducing discomfort and coincided with more creative hypotheses and a stronger sense of "talking to a person" (Abstract). - **Relevance:** (a) — | (b) live LLM structuring of discussion into shared maps (a visual-facilitation/recording function) and its side effects on attention | (c) AI structuring alone may suppress creative hypothesis generation in group problem-solving - **Caveats:** Abstract only; very small exploratory sample (8 dyads); no statistics in abstract. - **Screening check:** Meets criteria on abstract evidence — empirical, dyads interacting via an LLM-based mediation tool, method and n reported. #### OA1894 — A review of LLM-assisted ideation (61 studies) - **Citation:** Li, S., Padilla, S., Le Bras, P., Dong, J., & Chantler, M. (2025). A review of LLM-assisted ideation. arXiv:2503.00946 (v3, 27 Jul 2025). https://doi.org/10.48550/arXiv.2503.00946 - **Type:** review (systematic literature review, PRISMA) — peer-reviewed: no (preprint) - **Full text read:** yes (https://arxiv.org/pdf/2503.00946, v3) - **Method:** PRISMA-guided systematic review of Web of Science, IEEE Xplore, ACM DL, Scopus (+6 via Google Scholar); manual data extraction of domains, ideation stages, interaction forms, user-study methods; derives a 3-phase/7-stage "Hourglass Ideation Framework" (§2, pp. 3–5 of PDF). - **n:** 1,463 records → 1,024 after de-duplication → 78 after title/abstract screening → 61 included studies (§2.3, pp. 4–5 of PDF); 43 of 61 (70%) included a user study (§6, p. 29 of PDF). - **Setting:** LLM-assisted ideation research across design, writing, multimedia creation, scientific research and education; publications 2022–2024 (§7.1.1, p. 35 of PDF). - **Findings:** 1. 52 of 61 reviewed papers (85%) support individual ideation; only 9 (15%) are designed for group ideation; only "a few systems", e.g. Collaborative Canvas and Conversational Swarm Intelligence, support collaborative group processes (§5.3, p. 23 of PDF). 2. Stage imbalance: scope definition appears in 97% of studies but LLM involvement there is 5%; idea generation in 100% (75% using LLMs solely for direct idea generation); multi-idea evaluation and selection in 46% with LLM support in 14% (§7.1.2, p. 35 of PDF). 3. Named research gaps specific to groups: LLM support for "fostering effective group collaboration, ensuring transparent and equal participation, and facilitating consensus-building", free-riding detection, and structuring many ideas under group time constraints (§7.2, p. 37 of PDF). - **Relevance:** (a) — | (b) maps scarcity of LLM tools for group/facilitated ideation (9/61) | (c) evidence base on LLM ideation is mostly individual; convergent/selection phases under-studied - **Caveats:** preprint; review, not primary data; literature coverage ends around 2024; group-ideation subset is small. - **Screening check:** FAILS: secondary literature review, not an empirical primary study (useful as background/context and for citation chasing of its 9 group-ideation papers). #### C21 — Generative AI enhances individual creativity but reduces collective diversity - **Citation:** Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290. https://doi.org/10.1126/sciadv.adn5290 - **Type:** journal — peer-reviewed: yes - **Full text read:** yes (UCL Discovery accepted/published PDF, https://discovery.ucl.ac.uk/id/eprint/10195027/; open access CC BY) - **Method:** Preregistered two-phase online experiment: writers randomized to Human-only, Human + 1 GenAI idea, or Human + up to 5 GenAI ideas; stories rated blind by separate evaluators; story similarity measured via embedding cosine similarity (Introduction and Fig. 1, p. 1–2). - **n:** 293 writers (stories) and 600 evaluators (each rated 6 stories) (p. 1–2). - **Setting:** Online participants (Prolific-type platform; approval rating ≥95% criterion, Materials and Methods) writing eight-sentence short stories individually; optional three-sentence story ideas from OpenAI GPT-4 (p. 2). - **Findings:** 1. Access to GenAI ideas raised rated novelty by 5.4% (1 idea; b = 0.207, P = 0.021) and 8.1% (5 ideas; b = 0.311, P < 0.001), and usefulness by 3.7% (b = 0.185, P = 0.039) and 9.0% (b = 0.453, P < 0.001) respectively (Results, p. 3); gains concentrated among less creative writers (low DAT scores) (p. 3–4). 2. GenAI-enabled stories were more similar to the other stories in their condition (1 idea: b = 0.871, P < 0.001; 5 ideas: b = 0.718, P = 0.003), equal to 10.7% and 8.9% of the Human-only similarity range (p. 4); and 5.2% / 5.0% more similar to the GenAI ideas (P < 0.001) (p. 5). 3. Across the two GenAI conditions 88.4% of writers used at least one AI idea; 24.5% in the 5-idea condition requested all five (p. 2, p. 4). - **Relevance:** (a) — | (b) — | (c) indirect: individual AI-assisted ideation reduces collective diversity of outputs (homogenisation), relevant to AI use in group ideation, but no group interaction studied - **Caveats:** Individual writing task, not group work; online participants; outcomes are evaluator ratings of short fiction. - **Screening check:** FAILS: no ≥2 humans interacting as a group (individual writers; "collective" diversity is an aggregate across independent individuals), no facilitators as respondents, no facilitation function. Useful as background evidence for dimension (c) only. #### OA1450 — Collective creativity in human-AI hybrid networks (CogSci 2025) - **Citation:** Shiiku, S., Marjieh, R., Anglada-Tort, M., & Jacoby, N. (2025). The dynamics of collective creativity in human-AI hybrid societies. Presented at CogSci 2025 (Underline Science record https://doi.org/10.48448/qztd-bp85); full text: arXiv:2502.17962v3 (13 May 2025). Note: the batch metadata lists "Cognitive Science Society 2025" as an author — that is the Underline venue tag, not an author; the author list here is taken from the arXiv PDF. - **Type:** conference (talk record) / preprint — peer-reviewed: unclear (the arXiv version read is formatted as a CogSci paper; proceedings publication not verified) - **Full text read:** yes (https://arxiv.org/pdf/2502.17962v3; the Underline DOI itself only exposes the abstract, which matches the arXiv abstract) - **Method:** online network experiment: participants embedded in 5×5 grid social networks, each selecting and elaborating a neighbour's short story once, over 25 iterations; three conditions: human-only, AI-only (GPT-4o), human–AI (50/50; humans not told stories might be AI-generated); creativity rated 1–5 by a separate rater sample blind to source; diversity = inverse mean TF-IDF cosine similarity (Method, p. 3–4 of PDF). - **n:** 879 human participants (Prolific, UK, native English) + 996 GPT-4o calls; human-only N = 625, AI-only 625 API calls, human–AI N = 254 humans (50%) plus GPT-4o agents; 100 separate raters, 20 stories each (Method, p. 3–4 of PDF). - **Setting:** creative storytelling from a fixed seed story; GPT-4o (version of 3 Sept 2024); PsyNet platform; Max Planck Ethics Council approval (p. 3–4 of PDF). - **Findings:** 1. Creativity ratings: AI-only M = 3.571 (SD 1.026) was significantly higher than human-only M = 2.482 and human–AI M = 2.327 (p < .001); AI-only and human–AI networks improved over iterations (gains M = 0.464 and 0.264, p < .001), human-only did not (M = −0.133, p = 1.000) (Results, p. 4 of PDF). 2. Diversity: AI-only highest overall (M = 0.880), then human–AI (0.860), human-only (0.823); AI-only diversity declined over iterations (Δ M = −0.034) while human–AI rose most (Δ M = 0.098, 95% CI [0.064, 0.131]) and ended with the highest diversity in the final iterations (Results, p. 4 of PDF). 3. Semantic analysis: AI agents drifted from the seed story (e.g., converging on space themes), humans retained core elements ("John", "key", "cat"); in hybrid networks humans' stories interpolated between the two, which the authors read as humans being influenced by AI ideas (Results, pp. 4–5 of PDF). - **Relevance:** (a) — | (b) — | (c) mixing humans and LLM agents in a creative network sustained collective diversity better than AI-only over time, while AI-only scored highest on rated creativity - **Caveats:** indirect interaction (each person contributes once to a transmission network, no group discussion); simple storytelling task the authors describe as "relatively easy" for GPT (Discussion, p. 5); some reporting inconsistencies (human-only CI printed as [2.389, 2.576] while human–AI CI printed as [2.184, 2.576]; human-only vs human–AI described as "similar" with p < .001). Same lab as OA1723 (Marjieh, Jacoby) — related paradigm. - **Screening check:** borderline — empirical, n reported, humans and GPT-4o co-create in multi-person networks, but not a group interacting synchronously; include only if collective-network designs count as group settings. #### OA1723 — Human-AI synergy in collective creative search - **Citation:** Li, C., Marjieh, R., Hu, H., Steyvers, M., Collins, K. M., Sucholutsky, I., & Jacoby, N. (2026). Human-AI synergy supports collective creative search. arXiv:2602.10001 (v3, 14 Jul 2026). https://doi.org/10.48550/arXiv.2602.10001 - **Type:** preprint — peer-reviewed: no (preprint) - **Full text read:** yes (https://arxiv.org/pdf/2602.10001, v3) - **Method:** preregistration not stated; online experiment with transmission-chain networks in a word-guessing ("hidden target word") task; 4 conditions: Human Social, Human Asocial, AI-only, Human–AI Hybrid (rounds randomly assigned to human or AI; humans not told some guesses came from AI); performance = max Word2Vec similarity per round; lexical diversity; LLM-as-judge strategy annotation (Claude Sonnet 4); Bonferroni-corrected tests (Method, p. 3 of PDF; Fig. 3 caption, p. 5). - **n:** 503 Prolific participants (US, native English; mean age 38.13, SD 12.30); Human Social N = 210, Human Asocial N = 179, Hybrid N = 114; 200 games (10 target words × 5 games × 4 batches), each 10 rounds × 10 guesses; Hybrid had 1,058 human and 1,018 AI rounds; 165,140 Gemini 2.5 Flash queries, 10,020 GPT-5.1 queries (Method, p. 3 of PDF). - **Setting:** Gemini 2.5 Flash (temperature 0.7) as AI players; GPT-5.1 in control experiments; PsyNet platform; Cornell IRB (Method, p. 3 of PDF). - **Findings:** 1. Hybrid human–AI chains reached significantly higher similarity scores than all other conditions (p < .001); AI-only lowest and plateaued after round 5; improvement over rounds Δ = 22.10 for AI vs. 40.93–53.21 for other conditions (Results, p. 3–4 of PDF). 2. Lexical diversity: Human Asocial M = 0.86, Hybrid M = 0.58, AI-only M = 0.37 (Hybrid > AI-only, p < .001); Exploit strategies made up 99% of AI-only rounds and AI showed virtually no adaptation to hint quality (Results, p. 4 of PDF). 3. In the hybrid condition AI agents performed better than in AI-only chains (p < .001, d = 0.618), while humans did not differ from Human Social (p = 0.392, d = 0.032); a mixed Gemini + GPT-5.1 AI-only condition (M = 72.83) beat single-model AI chains (M = 65.25 and 64.08) but remained below Human–AI Hybrid (M = 87.76; t(4081) = −14.93, p < .001, d = 0.30) (Results, p. 5 of PDF). - **Relevance:** (a) — | (b) — | (c) mixing humans and LLM agents in a collective idea-search chain outperformed human-only and AI-only collectives; AI alone converges narrowly - **Caveats:** preprint; abstract word-guessing task, not real ideation or a workshop; humans interact only indirectly via passed-on best guesses (no synchronous group discussion); authors acknowledge the single-word task operationalisation as a limitation (Discussion, p. 6 of PDF). Likely overlaps with the same lab's hybrid-network work (OA1450, Shiiku et al.). - **Screening check:** borderline — empirical, n reported, generative AI embedded in multi-human collectives, but "group" is a transmission chain without direct interaction; include only if indirect collective-intelligence designs count as group settings. #### R1394 — GPT-4 as teacher in team co-ideation of serious games - **Citation:** Jost, P. (2024). AI-deation: When the Teacher is a Transformer in Role-Playing to create Privacy Decision Serious Games. International Journal of Serious Games, 11(2), 65-86. https://doi.org/10.17083/ijsg.v11i2.762 - **Type:** journal — peer-reviewed: yes - **Full text read:** yes (https://journal.seriousgamessociety.org/index.php/IJSG/article/download/762/540) - **Method:** between-group design; post-activity 7-point Likert survey; Mann-Whitney U between groups; one-sample Wilcoxon for AI-perception items (§2.3, §3, pp. 74-75) - **n:** 55 students: 32 in online teams with a human in the teacher role vs 23 in local teams with GPT-4 in the teacher role (§2.3.1, p. 74); number of teams not verified - **Setting:** Computer Science students at two European universities co-playing a card-based role-playing framework in two-hour sessions (one-hour co-playing) to ideate and balance privacy-decision serious-game concepts; GPT-4 plays the teacher role (§2.3, p. 74) - **Findings:** 1. Ideation support was largely comparable: "The boardgame provides guidance to develop new ideas" 70%/70% agreement (NoAI/AI), U = 349.5, p = 0.746 (Table 3, p. 75). 2. Only significant difference: "Cards helped to focus on ideas" 72% (NoAI) vs 57% (AI), U = 232, z = -2.392, p = 0.017, r = 0.37 (§3.1, p. 75). 3. No significant differences in balancing support (all p > 0.05) or satisfaction with the created concept (75% NoAI vs 78% AI, U = 387, p = 0.737) (Tables 4-5, pp. 76-77); the author identifies timing of scheduled AI interventions as a factor that can impair creative support (Abstract; §4, p. 18 of PDF). - **Relevance:** (a) — | (b) GPT-4 can take a facilitating/teacher role in a structured co-ideation session with roughly equivalent perceived support | (c) AI teacher slightly reduced perceived focusing of ideas; intervention timing matters - **Caveats:** conditions confounded with setting (online vs local) and university; student sample; perception measures only; single author - **Screening check:** meets criteria (education exception) — GPT-4 participates in a structured team ideation session and team-level ideation/balancing support is measured; method and n reported; English. #### R1497 — ChatGPT in collaborative software prototyping teams - **Citation:** Pires, A. V., Kleinpaul, E., Campos, L., & Parizi, R. (2025). From Prompts to Prototypes: Exploring ChatGPT in Collaborative Software Prototyping Education. Proceedings of the Brazilian Symposium on Software Quality (SBQS 2025), 453-462. https://doi.org/10.5753/sbqs.2025.13937 - **Type:** conference — peer-reviewed: yes (SBQS; not verified further) - **Full text read:** yes (https://sol.sbc.org.br/index.php/sbqs/article/download/39034/38806) - **Method:** case study; onboarding + hands-on prototyping activity with ChatGPT; individual and group (TAM-based) questionnaires, prompt analysis, direct observation with field notes; descriptive statistics and thematic analysis (§3, pp. 3-4 of PDF) - **n:** 25 undergraduate students in 4 groups (G1 = 6, G2 = 8, G3 = 5, G4 = 6) (§4.1.1 and Table 4, p. 4 of PDF) - **Setting:** Information Systems students at IFFar (Brazil) in a Software Engineering course grounded in Design Thinking; teams used ChatGPT to design and refine software prototypes during an in-class activity (§3, p. 3 of PDF) - **Findings:** 1. ChatGPT's influence on group decisions: low 28% (idea generation only), medium 64% (shaping the prototyping plan), high 8% (generating code/UI directly used) (§4.3, p. 5 of PDF). 2. 72% of participants reported blending ChatGPT's suggestions with their own ideas, 28% accepted AI output directly without major revisions; 64% believed results would have differed without the tool (§4.3, p. 5 of PDF). 3. Group questionnaire: "ChatGPT facilitated collaboration among group members" M = 4.50 (SD 0.58), whereas "ChatGPT helped generate better ideas than those initially proposed by the group" M = 2.50 (SD 1.29) and reliability M = 2.75 (Table 5, p. 6 of PDF); common pattern was one typist with others suggesting prompt content (§4.4, p. 5 of PDF). - **Relevance:** (a) — | (b) ChatGPT as a shared tool in a team design session; collaborative prompting patterns | (c) teams rarely felt AI produced better ideas than their own; high-influence use linked to less critical evaluation - **Caveats:** small single-class case study; descriptive only; student sample; group questionnaire n not explicit - **Screening check:** meets criteria (education exception) — generative AI used within a structured team prototyping (design-thinking) session and group-level decision-making/prompting processes are analysed; method and n reported; English. #### OA574 — Students' views of GenAI as co-creator in business model workshops - **Citation:** Petkov, K. (2026). Students' Perspectives on Generative AI as a Co-Creator in Business Model Workshops. Research Papers of UNWE (Научни трудове на УНСС), issue 5/2025 (DOI 2026.1), pp. 197–215. https://doi.org/10.37075/rp.2026.1.09 - **Type:** journal — peer-reviewed: unclear (journal states a peer-review policy; not verified for this article) - **Full text read:** yes (https://rp.unwe.bg/index.php/rp/article/download/50/42) - **Method:** mixed methods: post-workshop structured survey (MS Forms) + thematic analysis of open-ended answers; no control group (Methodology, pp. 201–202; Limitations) - **n:** 28 undergraduate students (p. 201) - **Setting:** University of National and World Economy, Sofia; business-model innovation workshop combining BMI Pattern Cards, brainstorming, ranking and evaluation, with students encouraged to use ChatGPT, Gemini and others; tutor facilitated and supplied sample prompts (pp. 201–202) - **Findings:** 1. 53.6% described AI as "Essential" for generating ideas and 46.4% as "Helpful, not critical"; none chose "Not helpful" or "No difference" (Results, p. 204). 2. 82.2% rated AI's improvement of idea quality at 4 or 5 on a 5-point scale; no ratings of 1 (p. 204). (Abstract rounds to 54% / 46% / 82%.) 3. 64.3% "Yes, definitely" and 35.7% "Maybe, in some aspects" supported future AI integration in workshops; 0% "No" (p. 206). - **Relevance:** (a) — | (b) GenAI used inside a facilitated group ideation workshop | (c) self-reported idea-quality improvement only - **Caveats:** student sample, n=28, self-report, no comparison group; the survey item on "group brainstorming" exists but group-level outcomes are not measured; perceived rather than rated idea quality. - **Screening check:** Borderline — empirical, n=28, GenAI in a structured group brainstorming workshop, but outcomes are individual student perceptions in a classroom course (EX-EDU risk: criteria require a group-level outcome or process to be measured, which is not the case). #### S2-511 — AI-assisted Business Model Canvas workshop for vocational students (Bali) - **Citation:** Putri, N. W. N. S. T., Suhendro, J. M., & Kusuma, N. P. N. (2026). Mindset in Vocational Students Through AI-Assisted Business Model Canvas Workshop. Lamahu: Jurnal Pengabdian Masyarakat Terintegrasi, 5(1), 139–146. https://doi.org/10.37905/ljpmt.v5i1.36640. The title is exactly as printed on p. 139; it reads as if a leading word may be missing, but this is not verified. The internal workshop title is "Workshop on the Application of the Business Model Canvas (BMC) Assisted by Artificial Intelligence (AI) to Enhance Technopreneurial Spirit among Students of SMK Saraswati 1 Denpasar" (p. 142). - **Type:** journal (community-service report) — peer-reviewed: unclear (received 6 Jan 2026, accepted 24 Jan 2026, published 2 Feb 2026, p. 139; the review process is not described) - **Full text read:** yes (local PDF pdf3/S2-511.pdf, 8 pages, https://doi.org/10.37905/ljpmt.v5i1.36640). **The article is written in English** (confirmed: abstract, body and tables are in English; only some references are Indonesian). - **Method:** Community-service intervention with descriptive evaluation and no control group. Stages were problem identification (interviews with the principal and teachers), needs analysis, one 2-hour workshop (4 Sep 2025) and evaluation. Instruments: a pre–post concept quiz (non-graded verbal Q&A), a BMC rubric applied to group outputs, and structured facilitator observations. Results are given as counts and percentages only (pp. 140–142; Table 1, p. 142). - **n:** 51 grade-12 students (Accounting, Digital Business, Office Administration majors) at SMK Saraswati 1 Denpasar, organised into **10 groups** (p. 140, §2; Table 2 header "n=51; 10 groups", p. 142). Group size is not stated (about 5 per group, implied). The number of facilitators is not stated ("Facilitators accompanied each group", p. 141). - **Setting:** Indonesia (Denpasar, Bali), vocational high school. Session 1 (30 min): concept talk. Session 2 (50 min): step-by-step BMC completion in groups with ChatGPT for ideation and refinement (value propositions, customer segments, channels, consistency checks), guided by a speaker, with facilitators mentoring each group and taking observation notes. Session 3 (40 min): group presentations with feedback (abstract, p. 139; pp. 141, 143). The ChatGPT version is not stated. - **Findings:** 1. Conceptual understanding: "Before the session, only 3.92% of students demonstrated adequate technopreneurship understanding… After the session, this increased to 100%" (p. 143; Table 2, p. 142). 2. Group outputs: "all groups completed all nine BMC blocks (100% completion)". Value Proposition "required improvement in 30% of groups", and "40% of groups confused Customer Relationships with Channels" (p. 143; Table 2, p. 142). 3. AI use: "facilitators observed an improvement in prompt specificity in 80% of groups". Early prompts were generic ("business ideas for students") and became more contextual with guidance; "AI outputs were selectively integrated into BMC refinement" (p. 144; Table 2–3, pp. 142, 144). Note the inconsistency: the abstract instead says "80% of group outputs reflecting technopreneurship-based business ideas" (p. 139), which the body does not report. - **Relevance:** (a) Facilitators and a speaker guided student groups' ChatGPT prompting within a structured workshop | (b) ChatGPT used as an in-session ideation and refinement aid in facilitated group work | (c) Descriptive only: all groups completed the BMC, with coherence weaknesses in 30–40% of groups; no comparison without AI. - **Caveats:** Weak descriptive design with no control. The pre–post quiz was a non-graded verbal Q&A delivered around the concept session, so its gain is not attributable to AI. Percentages across 10 groups are based on very small counts. Author-run evaluation. The abstract and body disagree on the 80% figure. Student sample. The authors note device and internet variability and no follow-up (p. 145). - **Screening check:** Borderline, but meets the criteria. It is empirical, with student groups working together with ChatGPT in a facilitated workshop, and it reports method and n (51 students, 10 groups). The article is in English. #### C74 — Chatbots and co-creation in collaborative holiday planning - **Citation:** Kuhail, M. A., Alimamy, S., Khan, S., & Mrabet, J. (2026). "Can Chatbots Foster Co-Creation in Collaborative Holiday Planning?" An Expectancy-Value Perspective. Information Systems Frontiers (published online 17 July 2026). https://doi.org/10.1007/s10796-026-10784-6 - **Type:** journal — peer-reviewed: yes - **Full text read:** abstract only (Springer landing page via WebFetch, where only the abstract and references were visible; no OA copy). The details below come from the landing-page abstract, reported through the WebFetch summary. - **Method:** Live multi-user chatroom study with an integrated AI chatbot, then a post-task survey. Reasons for and against chatbot use were modelled as predictors of attitudes and willingness to co-create (expectancy–value theory). The modelling technique (likely SEM) is not verified. - **n:** 112 participants in 56 pairs. - **Setting:** Pairs collaboratively planning a holiday in a chatroom with an "AI-powered chatbot". The specific model is not verified. - **Findings:** 1. Interaction quality, customisation, information quality and perspective-taking strengthened reasons for using chatbots, which improved attitudes and willingness to co-create. 2. Privacy concerns and low-quality advice did not significantly diminish attitudes. Personal values amplified reasons for use. (Exact coefficients not verified.) - **Relevance:** (a) — | (b) chatbot as a shared assistant in a dyadic planning conversation | (c) perceived chatbot qualities predict willingness to co-create; decision outcomes not reported - **Caveats:** Abstract only. Leisure planning in dyads, not a facilitated work session. Survey-model outcomes rather than behaviour. - **Screening check:** Borderline pass. It is empirical, 2 humans interact with a generative chatbot in a live chatroom, n is reported, and it is in English. Relevance to facilitation is weak. #### R778 — ChatGPT in collaborative travel planning (pairs) - **Citation:** Kumamoto, H., & Joho, H. (2025). Exploring the role of large language model in collaborative travel planning task. Information Research, 30(iConf), 1123-1130. https://doi.org/10.47989/ir30iconf46966 - **Type:** journal (iConference 2025 paper in Information Research) — peer-reviewed: yes - **Full text read:** abstract only (OA landing page publicera.kb.se blocked by bot protection) - **Method:** within-subject study; two conditions (Google only vs Google + ChatGPT); questionnaires, observation, conversation-volume and ChatGPT-log analysis; t-tests and Wilcoxon signed-rank tests (Abstract) - **n:** 20 participants in 10 pairs (Abstract) - **Setting:** pairs planning a 1-night, 2-day trip together; ChatGPT (model version not verified) - **Findings:** 1. No significant difference in task completion time between conditions (Abstract). 2. Plans made with ChatGPT were more complete and aligned with requirements (Abstract; effect sizes not verified). 3. Participants found information more easily with ChatGPT, but satisfaction levels remained similar (Abstract). - **Relevance:** (a) — | (b) — | (c) genAI as shared tool in a two-person joint planning/decision task: better plan quality, no efficiency or satisfaction gain - **Caveats:** abstract only; very small n; dyadic (2-human) groups; the task is a joint information-search/planning task rather than a facilitated session - **Screening check:** meets criteria — empirical within-subject study, 2 humans collaborating as a group with ChatGPT, method and n reported, English. ## 4. AI for planning, running and following up meetings #### OA360 — MeetMap: real-time collaborative dialogue mapping with LLMs - **Citation:** Chen, X., Yap, N., Lu, X., Gunal, A. E., & Wang, X. (2025). MeetMap: Real-Time Collaborative Dialogue Mapping with LLMs in Online Meetings. Proceedings of the ACM on Human-Computer Interaction, 9(2), Article CSCW132, 1–35. https://doi.org/10.1145/3711030 - **Type:** journal (PACM HCI, CSCW) — peer-reviewed: yes - **Full text read:** yes (arXiv:2502.01564v1, formatted as the PACM HCI article; https://arxiv.org/pdf/2502.01564; page numbers = CSCW132:N) - **Method:** Within-subject lab experiment, three conditions per pair: Human-Map (AI summarises speech into nodes, users build the map), AI-Map (AI drafts the dialogue map, users edit), baseline (Zoom + shared Google Doc with AI-generated summaries); 30-min decision discussion per condition; surveys, log data, interviews, video memos (§4, pp. 3, 16). - **n:** 20 participants in 10 pairs (p. 3). - **Setting:** Online video meetings with decision-making tasks; GPT-4 API used for IBIS tagging, summaries and map generation (§3, pp. 9–13). - **Findings:** 1. Both MeetMap variants were rated higher than baseline for keeping up with the discussion (reported as p = 0.017 and "p = 0.46 < 0.05" — the second value appears to be a typo in the paper), for reflecting the decision-making process (Human-Map p = 0.001; AI-Map p = 0.003) and for summary accuracy (p = 0.03; p = 0.002) (§5.1.1, p. 16). 2. Users rated Human-Map (p = 0.003) and AI-Map (p = 0.004) significantly higher than baseline for helping the team reach consensus; 8 of 20 said MeetMap built shared understanding, and AI-generated nodes served as "objective mediators" for resolving misunderstandings (§5.1.2, p. 17). 3. All 20 users found the full transcript in the baseline excessive and preferred the concise AI nodes; Human-Map gave more sense of control and agency, while users tolerated AI errors less when they had invested effort in the map (pp. 16–17). - **Relevance:** (a) — | (b) LLM support during meetings (real-time visual structuring of discussion) | (c) perceived consensus support and shared understanding (self-report) - **Caveats:** Pairs only; 10 pairs; outcomes self-reported, no objective decision-quality measure; test statistics not fully reported in the text. - **Screening check:** Meets criteria (empirical within-subject study, pairs of humans interacting with LLM meeting support, method and n reported). #### OA355 — CoExplorer: GenAI adaptive interface for planning and running video meetings - **Citation:** Park, G. W., Panda, P., Tankelevitch, L., & Rintel, S. (2024). The CoExplorer Technology Probe: A Generative AI-Powered Adaptive Interface to Support Intentionality in Planning and Running Video Meetings. In Proceedings of the 2024 ACM Designing Interactive Systems Conference (DIS '24), pp. 1638–1657. https://doi.org/10.1145/3643834.3661507 - **Type:** conference — peer-reviewed: yes - **Full text read:** yes (arXiv:2405.18239v2, https://arxiv.org/pdf/2405.18239; page numbers = preprint pages) - **Method:** Technology-probe study: 26 one-hour sessions, each with one participant and a researcher, in a controlled simulation of the meeting lifecycle (other attendees simulated with pre-assigned opinions); semi-structured interview and think-aloud; closed and open coding (§4, pp. 7–8). - **n:** 26 employees of one global technology company (18 male, 8 female; 31% project managers, 31% software engineers, 23% researchers, 12% designers, 4% hardware engineers) (§4.1, p. 8). - **Setting:** Prototype that generates meeting phases from an invitation, a pre-meeting "Meeting Focus Tool" to capture attendee preferences, phase-based app/window layouts, and conversation-based phase refinement; GPT-3.5 and GPT-4 (§3, p. 5); ChatGPT used to pre-generate study materials (§4.2). - **Findings:** 1. 24 of 26 participants found capturing attendees' preferences in advance helpful; 17 of 26 wanted it well before the meeting as preparation; 14 of 26 appreciated the automatic "democratisation" of meeting flow (§5.2, p. 9). 2. 14 of 26 wanted phase-refinement results to go through the meeting organiser; 11 of 26 worried the phase list could become too focused or overwrite the organiser's agenda; six opposed pre-meeting and eight during-meeting refinements (§5.2.1–5.2.2, p. 9). 3. 21 of 26 said the system should consult humans before showing any file/app in the shared space and 16 of 26 wanted checks before phase changes (§5.4.1, pp. 10–11). - **Relevance:** (a) meeting organisers' attitudes to AI-generated agendas/phases (participants were employees, not specifically facilitators) | (b) GenAI planning meeting structure and adapting it live; strong desire for organiser control | (c) — - **Caveats:** Single company; simulated meeting with one real participant; researcher-triggered phases; attitudes to a probe, not actual use. - **Screening check:** FAILS (strict): EX-INDIV — each session had a single participant with simulated attendees, so no group of ≥2 humans interacted with the AI, and participants were general employees rather than facilitators reporting their own AI use. Borderline: highly relevant to AI-supported meeting planning; the lead reviewer may keep it as contextual evidence. #### C0 — Are We On Track? AI-assisted goal reflection during meetings - **Citation:** Chen, X., Tankelevitch, L., Vanukuru, R., Scott, A. E., Panda, P., & Rintel, S. (2025). Are We On Track? AI-Assisted Active and Passive Goal Reflection During Meetings. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI '25), 22 pages. https://doi.org/10.1145/3706598.3714052 - **Type:** conference (full paper) — peer-reviewed: yes - **Full text read:** yes (arXiv version, https://arxiv.org/pdf/2504.01082v2) - **Method:** Technology probe study with video-stimulated recall and semi-structured interviews; each participant's own real meeting recording/transcript was loaded into two GPT-4-based probes (passive "Ambient Visualization" vs. active "Interactive Questioning"); thematic analysis (§3, p. 4–7 of PDF). - **n:** 15 knowledge workers from one global technology company; 12 real meeting recordings (meeting sizes 2–24 attendees, 30 min–1 h) (§3.1 p. 4; Table 2 p. 8; footnote 4: "fewer meetings (n = 12) than participants (n = 15)"). - **Setting:** Employees of a global technology company; real workplace meetings (stand-ups, team syncs, project meetings, onboarding) replayed in a simulated meeting interface; GPT-4 detects topic changes, summarizes, infers goals (passive probe) or decides timing and generates reflection questions that attendees vote on (active probe) (§3.2, p. 6). - **Findings:** 1. Goal clarification was the primary value for most participants for both probes (11/15); with the active probe participants frequently began by asking "What are we trying to achieve?" (13/15); with the passive probe most (10/15) first interpreted AI-generated goals rather than generating their own (§5.1.1, p. 9). 2. Benefits: both probes helped surface/clarify meeting objectives (9/15) and encouraged ownership of participation (12/15); AI reflection was seen to foster shared responsibility (9/15), with AI acting as a "neutral mediator" so that nobody had to interrupt personally (§5.2.1, §5.2.3, p. 10–11). 3. Concerns: passive visualization risked cognitive overload (5/15) and was often ignored (6/15); active questions were sometimes too generic/irrelevant (6/15), felt intrusive (4/15) or created pressure to respond (5/15); with the passive probe 6 participants initially accepted AI-inferred goals as accurate before spotting discrepancies (over-reliance) (§5.3.1–5.3.3, p. 11). - **Relevance:** (a) — | (b) AI (GPT-4) monitoring goal alignment and prompting group reflection in real meetings, a core in-meeting facilitation function; design trade-offs of intervention intensity/timing | (c) indirect: effects on meeting focus/prioritisation, not decision quality measured - **Caveats:** Probes not deployed live; counterfactual reactions to replayed own meetings (authors acknowledge this, p. 6); single company (Microsoft Research authors); qualitative counts only. - **Screening check:** Meets criteria — empirical, n reported, LLM (GPT-4) performing an in-meeting facilitation function evaluated with meeting participants/organizers on real group meetings. Borderline: participants were interviewed individually, not interacting live as a group with the AI. #### C8 — What Does Success Look Like? AI-assisted prospective meeting reflection - **Citation:** Scott, A. E., Tankelevitch, L., Panda, P., Vanukuru, R., Chen, X., & Rintel, S. (2025). What Does Success Look Like? Catalyzing Meeting Intentionality with AI-Assisted Prospective Reflection. In Proceedings of the 4th Annual Symposium on Human-Computer Interaction for Work (CHIWORK '25), 32 pages. https://doi.org/10.1145/3729176.3729204 - **Type:** conference (full paper) — peer-reviewed: yes - **Full text read:** yes (arXiv version, https://arxiv.org/pdf/2505.14370v1) - **Method:** Qualitative technology-probe study using a "participatory prompting" methodology: onboarding survey (3 upcoming meetings per participant), ~60-min moderated session with the Meeting Purpose Assistant (3 × 10 min conversations + 5 min discussion of each Reflection Summary), follow-up survey after the meetings took place (§4.3, p. 5–7 of PDF). - **n:** 18 employees (19 recruited, 1 excluded; 10 women, 8 men) from six work areas of one global technology company; up to 3 upcoming meetings each (§4.2, p. 5). - **Setting:** Meeting organizers/attendees reflecting individually, before meetings, on the purpose, success criteria and challenges of their own upcoming meetings with a chat assistant powered by GPT-4 Turbo (OpenAI Assistants API), with meta-prompts varied for organizers vs. attendees; the assistant produces a "Reflection Summary" (§4.1, p. 5; Fig. 2, p. 6). - **Findings:** 1. Observed impacts: clarifying meeting purposes, challenges and success conditions; changing perspectives; improving preparation and communication; proposing changed plans (Abstract, p. 1). Several participants said they would paste the AI Reflection Summary into meeting descriptions or chat threads (P10, P12, P14, P19) or share goals at the meeting start, and a subset reported in the follow-up survey having actually used it (§5.1.3, p. 9). 2. In the follow-up survey several participants reported the MPA interaction helped their meeting stay focused and on track (P1, P2, P4, P9); participants also saw potential to eliminate unnecessary meetings or their own attendance (§5.3.3, p. 11). Counts are named participants only; no quantitative effect measured. 3. Barriers: some found the interaction generic or "high-level" (P6, P7, P8, P16, P18), partly because recurring-meeting topics are set at short notice; solution-oriented participants (7 named) wanted the MPA to propose concrete improvements and questioned the value of reflection-only questioning (§5.4.1, §5.4.3, p. 12). - **Relevance:** (a) — (meeting organizers, not professional facilitators) | (b) LLM coaching for meeting planning/preparation (purpose, success criteria, challenges) — pre-session design support | (c) — - **Caveats:** Individual use, not group interaction; researcher moderated the sessions and sometimes helped interpret MPA output (§4.3.2, p. 7); single company; same Microsoft Research team as C0. - **Screening check:** Borderline. Empirical, method and n reported, LLM (GPT-4 Turbo) performs a meeting-planning function; however respondents are general employees (organizers/attendees) acting individually, not ≥2 humans interacting as a group with the AI and not facilitators. Include only if "AI support for meeting planning by organizers" counts under (b); otherwise FAILS: no group interaction and respondents are not facilitators/meeting professionals. #### OA361 — LLM-powered meeting recap system (highlights vs. hierarchical minutes) - **Citation:** Asthana, S., Hilleli, S., He, P., & Halfaker, A. (2025). Summaries, Highlights, and Action Items: Design, Implementation and Evaluation of an LLM-powered Meeting Recap System. Proceedings of the ACM on Human-Computer Interaction, 9(2), Article CSCW176, 29 pp. https://doi.org/10.1145/3711074 - **Type:** journal (PACM HCI / CSCW) — peer-reviewed: yes - **Full text read:** yes (arXiv author version 2307.15793v3, https://arxiv.org/pdf/2307.15793v3; page numbers refer to the printed CSCW176:N pagination in that version) - **Method:** design + implementation of two recap prototypes (highlights; hierarchical "chapters"); preliminary task-based evaluation via semi-structured interviews with think-aloud, each participant using a transcript of one of their own recent work meetings (§4) - **n:** 7 information workers at Microsoft (§4.1, p. 11); each brought one real meeting recording 1–2 weeks old (§4.2, p. 11) - **Setting:** Microsoft employees' real work meetings; recaps generated by fine-tuned transformer models (deBERTa and BART for extraction/abstraction/segmentation/titles, §3.2, pp. 7–9) — the paper labels them "LLM-powered"; not a GPT-style chat LLM - **Findings:** 1. Highlights recap: four participants used suggested summaries as reminders; six said some generated highlights were not relevant to them (inadequate personalization) (§5.2.1, p. 14). 2. Group work: five participants wanted to edit notes collaboratively, seeing shared notes as a way to build consensus and transparency on tasks (§5.2.2, p. 16); with the hierarchical recap five found it easy to copy-paste discussion chapters to share with colleagues, but none said they would use it to share action items/decisions for consensus or accountability (§5.3.2, pp. 19–20). 3. Hierarchical recap: six participants said a quick overview was useful (§5.3.1, p. 18); user edits (adding/editing) were consistent in intent, deletions were inconsistent (done when task complete, redundant or inaccurate), limiting their use as training signal (§5.2.3, p. 17). - **Relevance:** (a) — | (b) AI post-meeting support (recap/minutes/action items) for real organisational meetings | (c) indirect: recaps seen as consensus-building artifact, no decision-quality measure - **Caveats:** n=7, individual interviews (not group sessions); models are fine-tuned BART/deBERTa, not a general-purpose generative LLM; Microsoft-internal authors and sample. - **Screening check:** Borderline pass. Empirical, method and n reported, generative (seq2seq) summarisation of real multi-person meetings; but users evaluated the recap individually, and "LLM" is used loosely. Could be judged EX-INDIV by a strict reader. #### OA396 — Togedule: LLM-adaptive group meeting scheduling - **Citation:** Song, J., Ashktorab, Z., & Malone, T. W. (2025). Togedule: Scheduling Meetings with Large Language Models and Adaptive Representations of Group Availability. Proceedings of the ACM on Human-Computer Interaction, 9(7), Article CSCW332, 24 pp. https://doi.org/10.1145/3757513 - **Type:** journal (PACM HCI / CSCW) — peer-reviewed: yes - **Full text read:** yes (arXiv 2505.01000v5, https://arxiv.org/pdf/2505.01000v5; pages = printed CSCW332:N) - **Method:** formative study, then two controlled within-subjects experiments (attendees; organizers) comparing Words (verbal messages) vs. Calendar (shared calendar) vs. Togedule, order counterbalanced (§4, §6) - **n:** formative N=10 via Upwork (6 attendees, 4 organizers) (§4.1, p. 7); experiments N=66 via Prolific: 48 attendees in 6 groups (3×6, 3×10) and 18 organizers, each organizer assessing 2 groups × 3 conditions (§6.2, p. 14) - **Setting:** asynchronous group meeting scheduling with artificial calendars; GPT-4 ("gpt-4-turbo-2024-04-09", temperature 0.1) rates voting status to decide how many time options to show next (§5, pp. 12–13) - **Findings:** 1. Organizer decision quality: 72.22% of decisions rated good with Togedule vs. 22.22% (Calendar) and 61.11% (Words); Togedule vs. Calendar significant (p < 0.01, t = 2.965, d = 0.699), vs. Words not significant (§7.3.1, p. 19). 2. Organizers were faster with Togedule than Calendar (p < 0.05, t = −2.606, d = 0.614) and reported lower mental load (p < 0.01, t = 4.610, d = 1.087) and less effort (p < 0.01, t = 5.169, d = 1.218) than with Words (§7.3.2–7.3.3, pp. 19–20). 3. Attendees: only mental load was significantly lower than Words (p < 0.05, t = 2.483, d = 0.358); time not different (e.g., vs. Words p = 0.749, d = 0.047); 23 of 48 preferred Togedule, 17 Calendar, 8 Words (§7.2.3–7.2.5, pp. 18–19). - **Relevance:** (a) — | (b) LLM supporting meeting logistics (scheduling) for groups — pre-session planning, not facilitation of the session itself | (c) organiser scheduling decision quality - **Caveats:** crowdworkers with artificial schedules; asynchronous, no real group interaction; decision quality rated by one professional scheduling assistant. - **Screening check:** Borderline — empirical, method and n reported, GPT-4 involved, multi-person groups; but participants do not interact as a group and the task is scheduling rather than facilitation (possible EX-TOPIC). Include only if meeting-planning support is in scope. #### S2-190 — Personalized jargon support for real-time online meetings (ParseJargon) - **Citation:** Song, Y., Liu, Y., Au, W. Y., Wong, H. Y., Bailey, B. P., & August, T. (2026). Breaking the Curse of Knowledge: Designing Personalized Jargon Support for Real-Time Online Meetings. arXiv preprint arXiv:2508.10239v3 [cs.HC] (v3 dated 7 Jun 2026; v1 2025). https://arxiv.org/abs/2508.10239 — duplicate versions of the same work: S2-192 (CHI EA 2026 version) and S2-203 (earlier arXiv version); count once. - **Type:** preprint (v3; a CHI EA 2026 version exists as S2-192) — peer-reviewed: no (this version) - **Full text read:** yes (arXiv 2508.10239v3 PDF, pdf3/S2-190.pdf; page refs = p. X of PDF, no printed page numbers) - **Method:** Three parts: (1) controlled within-subjects lab study (3 conditions: no support / generic LLM jargon support / personalized support filtered by a one-sentence profile), Wilcoxon signed-rank with Holm–Bonferroni; participants watched each other's *recorded* presentations individually, plus follow-up interviews (§3); (2) simulation-based system evaluation of refined personalization (in-session feedback, cross-session glossary) and backend latency on the controlled-study data (§5); (3) single live pilot deployment in a real team Zoom meeting without comparison condition, post-meeting survey incl. NASA-TLX (§6). - **n:** Controlled study: 7 participants (interns, one technology company), 7×6 = 42 participant-presentation pairs, 14 per condition (p. 1, §3.2). Latency test: 787 caption chunks across 7 recordings (1,574 generic / 3,148 personalized chunk measurements over two runs) (p. 8, §5.3). Pilot deployment: 10 members of one weekly team meeting (1 director, 6 senior, 3 junior researchers), one ~45-min session, no participant overlap with the controlled study (p. 8, §6.1.1–6.1.2). - **Setting:** Technology company (affiliations: University of Illinois Urbana-Champaign, Fujitsu Research of America). LLM = gpt-4o-2024-05-13 for identifying, defining and personalized filtering of jargon from meeting-platform captions, shown as a glossary sidebar (p. 3, footnote 2). Controlled study used Microsoft Teams recordings with manually verified transcripts (§3.2.2); deployment as a Chrome extension in a Zoom meeting with three 10-min junior-researcher presentations plus short discussions (§6.1.2). - **Findings:** 1. Personalization cut glossary size by 57% (22.57 → 9.71 terms per participant-presentation) and raised the helpfulness rate (precision) from 47.03% (generic) to 77.51% (personalized) (p. 4, Table 1, §3.3). Self-reported comprehension confidence: baseline 3.07±0.62, generic 3.79±0.70 (sig. vs baseline), personalized 4.29±0.61 (sig. vs baseline and generic). Engagement: 3.93 / 3.64 / 4.29 (generic lower than baseline; personalized not significant vs baseline). Perceived value: 3.57 / 3.93 / 4.43 (personalized sig. vs baseline and generic); corrected p<.05, Wilcoxon (p. 5, Table 2). 6 participants said the generic condition overwhelmed them with too many definitions (p. 5, §3.3). 2. Simulation: in-session feedback raised helpfulness from 77.51% to 83.41%, cross-session reuse to 80.79%, both combined to 83.95% (p. 7–8, Table 3, §5.2). Backend latency median/95th percentile 0.32/1.36 s (generic) vs 0.42/1.73 s (personalized), below the median caption display time of 4.15 s (p. 8, Table 4, §5.3). 3. Live 10-person Zoom deployment: usability ratings were ease of use 4.3 (SD 0.82), feature integration 4.5 (SD 0.53), willingness for future use 3.9 (SD 1.1). NASA-TLX was low (e.g., mental demand 4.8±3.39 of 21). Perceived effect on comprehension 3.9 (SD 1.37), engagement 3.6 (SD 1.35), value of colleagues' work 4.0 (SD 1.05). Two participants reported low willingness and usefulness because of jargon-identification inaccuracies, which appeared to stem mostly from caption errors. The within-team meeting also had little unfamiliar jargon (p. 9, §6.2.1–6.2.3, Table 5). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: LLM gives in-meeting comprehension support to individual attendees of a live multi-person meeting (one 10-person deployment), not facilitation of the group process | (c) AI and group ideation/decisions: — (presenter-graded takeaways/questions in the recorded-presentation study; no group decision measures) - **Caveats:** The controlled study (n=7) is individual viewing of recorded presentations, not a live group interaction. The only live group use is one uncontrolled pilot meeting (n=10) with no comparison condition; the authors say it was not designed to measure effectiveness (§6). Everything comes from a single company. The system evaluation is simulation-based (§8). This is a preprint; S2-192 and S2-203 are versions of the same work. - **Screening check:** Borderline pass. The paper is empirical, reports its methods and n, and is in English. It qualifies only through the §6 live deployment, where 10 people interacted in a real Zoom meeting while using GPT-4o meeting support; the controlled study alone would not meet the ≥2-humans-as-a-group criterion. The AI supports individual comprehension rather than facilitation, so it is weakly relevant to the review question. #### OA1539 — LADICA: LLM-assisted shared display for co-located teams (CHI 2025) - **Citation:** Zhang, Z., Peng, W., Chen, X., Cao, L., & Li, T. J.-J. (2025). LADICA: A large shared display interface for generative AI cognitive assistance in co-located team collaboration. In CHI Conference on Human Factors in Computing Systems (CHI '25), Yokohama. ACM, 22 pp. https://doi.org/10.1145/3706598.3713289 (preprint: arXiv:2409.13968v2) - **Type:** conference — peer-reviewed: yes (the arXiv v2 read here is the CHI '25 camera-ready; the batch metadata lists it as an arXiv preprint) - **Full text read:** yes (https://arxiv.org/pdf/2409.13968, v2, 7 Feb 2025) - **Method:** system design + formative study (focus group and scenario-based design workshop) + in-person lab user study without a control condition: post-study questionnaire, interaction-log/video analysis, semi-structured interviews with reflexive thematic analysis (§3, §5; PDF pp. 5, 14–17). - **n:** formative study: 8 participants in 2 sessions of 4 (§3.1, PDF p. 5); user study: 14 participants (5 male, 9 female; mean age 23.9) in groups of 3–4 (number of groups not stated explicitly; "Group 1" and "Group 2" named), 65-min sessions with two 15-min tasks (§5.1–5.3, PDF pp. 14–15). - **Setting:** university students (9 undergraduate, 5 graduate), all with LLM experience; 65-inch multi-touch display plus one iPad per participant; tasks: group spring-break trip planning and a flipped-classroom policy discussion; AI features (goal decomposition, idea expansion, affinity lenses, relation hints, speech-based key-information extraction) built on OpenAI GPT-4 turbo (§4.4, §5.2–5.3; PDF pp. 13–15). - **Findings:** 1. Feature use across sessions: group goal decomposition 7 times; query-based idea expansion 18 and relation-based expansion 15 times — in 19 of these 33 instances the individual sought input from other members before deciding which subtask to explore; in 3 instances participants dismissed AI-suggested subtasks as too generic or trivial (§5.4.2, PDF p. 16). 2. Group-based ideation hints were activated 12 times: suggested ideas were adopted directly in 5 instances and used as inspiration for participants' own ideas in 7; relation hints used 10 times, but some participants stopped using them because they cluttered the interface; one group abandoned the affinity-lens feature after one try, preferring to find themes themselves "to develop mutual understanding" (§5.4.2, PDF p. 16). 3. Interviews: LADICA was seen as an inspiration source in individual phases and "a facilitator for identifying differences or similarities in ideas" in collective analysis; concerns included off-topic/general suggestions (P4, P6, P7), clutter (P2, P6, P9), and that low-effort note creation might let "less diligent teammates … appear engaged while spending less time actively thinking" (P8) (§5.4.3, PDF pp. 16–17). Questionnaire results are reported only graphically (Fig. 11), so no numeric ratings can be cited. - **Relevance:** (a) — | (b) LLM-powered shared whiteboard supporting in-person team ideation, organisation and discussion (facilitator-like functions) | (c) qualitative evidence on how groups take up or reject AI ideas; no outcome comparison - **Caveats:** small student sample, no control/baseline, short sessions, single author coded interaction logs, questionnaire only as a figure; authors themselves note college-student-only sample (§7, PDF p. 19). - **Screening check:** meets criteria — empirical lab study, n reported, groups of 3–4 humans interacting with an LLM-based system. #### C32 — AI as an agent and collaborative space in small-group sync/async collaboration - **Citation:** Xu, H., Kharrufa, A., Solaiman, E., & Vlachokyriakos, V. (2026). AI as an Agent and Collaborative Space: Exploring the role of Generative AI in Small Group Synchronous and Asynchronous Collaborative Dynamics. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26), pp. 1–17. https://doi.org/10.1145/3772318.3791087 - **Type:** conference (full paper) — peer-reviewed: yes - **Full text read:** abstract only (ACM DL blocked; no preprint found) - **Method:** Qualitative study: observations and semi-structured interviews, analysed through the lens of distributed cognition (Abstract). - **n:** 27 higher-education students (number of groups not stated in abstract — not verified). - **Setting:** Small student groups using generative AI in synchronous and asynchronous teamwork; tools/models not stated in abstract (not verified). - **Findings:** 1. In synchronous settings, shared use of GenAI supported transparency and mutual awareness, with the interaction space acting as attentional anchor, shared memory and negotiable contributions to group decisions (Abstract). 2. In asynchronous teamwork, GenAI was typically used individually and outputs were later introduced into discussions, reducing opportunities for negotiation (Abstract). - **Relevance:** (a) — | (b) shared (projected/common) GenAI use during live group work supports group awareness and decision negotiation | (c) individual pre-use of GenAI reduces negotiation in group decisions - **Caveats:** Abstract only; student sample; qualitative. - **Screening check:** Meets criteria on abstract evidence — empirical, small groups interacting with generative AI, method and n reported. ## 5. Large-scale AI-mediated deliberation platforms (Conversational Swarm Intelligence, civic synthesis) #### OA2022 — CSI enhances groupwise deliberation (Thinkscape, 48-person groups) - **Citation:** Rosenberg, L., Willcox, G., Schumann, H., & Mani, G. (2024). Conversational Swarm Intelligence (CSI) Enhances Groupwise Deliberation. In Proceedings of the International Joint Conference on Advances in Computational Intelligence (IJCACI 2023), Algorithms for Intelligent Systems, Springer, pp. 1–14. https://doi.org/10.1007/978-981-97-0180-3_1 (preprint: arXiv:2309.12366) - **Type:** conference — peer-reviewed: unclear (Springer proceedings; the version read is the authors' arXiv preprint) - **Full text read:** yes (https://arxiv.org/pdf/2309.12366, the preprint; the Springer version is paywalled and was not compared) - **Method:** Within-group crossover comparison: 2 sessions, each group used standard chat for one question and CSI for the other, with the order reversed between sessions; paired t-tests on contribution rates; binomial tests on a post-session preference survey (Session 2 only) (p. 6–10, §3–4). - **n:** 2 separate groups of 48 paid participants from a professional sample provider (96 total, no one in both sessions); survey = Session 2 participants (p. 6, 10). - **Setting:** Online text chat on the Thinkscape platform (Unanimous AI). CSI splits the group into small rooms linked by LLM "Observer/Surrogate" agents that relay summaries between rooms; GPT-3.5 labels preferences. Questions: which jobs are most at risk from AI, and the biggest AI risk in 10 years; 6 minutes per question (p. 4–7). - **Findings:** 1. With CSI, participants sent 46% more messages (0.699 vs 1.021 messages/min, p<0.001) and 51% more characters (32 vs 49 chars/min, p<0.001) than in standard chat (p. 7, §4). 2. Participation was more balanced: the 90th/10th-percentile contribution ratio was 13% higher in standard chat by messages/min and 37% higher by characters/min (p. 9, §4). 3. Survey: 65% preferred CSI (18% no preference, 16% opposed; p<0.02); 67% felt more heard in CSI (24%/8%; p<0.005); on "better justifications" 51% favoured CSI (29%/20%; reported as "p<0.5") (p. 10, §4). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: LLM agents link parallel small groups and aggregate preferences, a scaled-up facilitation infrastructure | (c) AI and group ideation/decisions: more contributions and more balanced participation; decision quality not measured - **Caveats:** Vendor-authored (Unanimous AI). Paid online panel; only 2 sessions; the survey covered only Session 2. The abstract and conclusion cite "more impactful (p<0.01)", but the results section reports "felt more heard (p<0.005)" (wording/p-value mismatch). The p<0.5 for justifications is as printed. The same result is summarised in OA2071, OA2121 and OA2540 (same author group). - **Screening check:** PASSES (empirical; human groups interacting with LLM agents; method and n reported). #### OA2071 — CSI enables rapid group insights (81 US Republican voters) - **Citation:** Rosenberg, L., Willcox, G., & Schumann, H. (2023). Conversational Swarm Intelligence (CSI) Enables Rapid Group Insights. 2023 IEEE 14th Annual Ubiquitous Computing, Electronics & Mobile Communication Conference (UEMCON), pp. 0534–0539. https://doi.org/10.1109/uemcon59035.2023.10316130 (preprint: arXiv:2310.13702) - **Type:** conference — peer-reviewed: unclear (IEEE conference; the version read is the authors' preprint) - **Full text read:** yes (https://arxiv.org/pdf/2310.13702, preprint marked for UEMCON 2023) - **Method:** Single-condition pilot deployment (no control group); LLM-based preference labelling of chat; paired t-tests on candidate sentiment across three time periods (p. 3–5, §III–IV). - **n:** 81 US Republican voters in one session, randomly split into 15 parallel chat rooms of about 5–6 people (p. 3, §III). - **Setting:** Online text chat on Thinkscape. Task: forecast which of 6 Republican primary candidates the US public would most support, and argue reasons, in 6 minutes. An LLM agent in each room relays summaries between rooms; GPT-3.5 labels preferences every 5 messages or 15 s (p. 3–4). - **Findings:** 1. In 6 minutes the group converged on one candidate (DeSantis) and voiced 410 reasons (266 for, 144 against), an average of 5 per participant; 206 reasons were for DeSantis and 54 against (p. 4–5, §IV, Table I). 2. About 60 of 81 users (74%) backed DeSantis at the start and 55 (68%) at the end, so few switched. His average sentiment was significantly higher than the other candidates' in every time period (p<0.01; abstract: p<0.001) (p. 5, §IV, Table II). 3. The system output verbatim LLM summaries of arguments for and against each candidate, e.g., "62 participants argued in favor" of DeSantis and "24 participants argued against" (p. 4, §IV). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: LLM agents run cross-room information flow and summarise arguments in a large online deliberation | (c) AI and group ideation/decisions: fast convergence and argument surfacing, but little opinion change and no accuracy benchmark - **Caveats:** Vendor-authored; no control condition; "consensus" largely reflects the initial majority (74% at start); GPT-3.5 labelling validated only as "deemed highly accurate" without figures; forecast accuracy not assessed; the body gives p<0.01, the abstract p<0.001. - **Screening check:** PASSES (empirical; humans interacting in groups with LLM agents; method and n reported) — weak design. #### OA2121 — Towards Collective Superintelligence: CSI estimation pilot (241 participants) - **Citation:** Rosenberg, L., Willcox, G., & Schumann, H. (2023). Towards Collective Superintelligence, a pilot study. 2023 International Conference on Human-Centered Cognitive Systems (HCCS), pp. 1–6. https://doi.org/10.1109/hccs59561.2023.10452485 (preprint: arXiv:2311.00728) - **Type:** conference — peer-reviewed: unclear (IEEE conference; the version read is the authors' preprint) - **Full text read:** yes (https://arxiv.org/pdf/2311.00728, preprint) - **Method:** Pilot comparison of estimation accuracy: individual survey vs survey mean (wisdom of crowds) vs CSI conversational deliberation vs GPT-4 alone; one-tailed z-test for CSI vs average individual (p. 3–5, §IV–V). - **n:** 241 real-time participants; the CSI condition used 47 subgroups of 5–6 members (p. 3–4, §IV). - **Setting:** Online Thinkscape text chat; estimating the number of gumballs in a jar from a photo, choosing among 10 answer options within 4 minutes. An LLM agent in each subgroup passes top choices and rationales to a neighbouring subgroup. GPT-4 was also given the photo (p. 4–5). - **Findings:** 1. Mean absolute error: individuals 361 gumballs (55%); GPT-4 279 (42%); survey mean 163 (25%); CSI conversational estimate 82 (12%) (p. 5, §V). 2. CSI error was 77% smaller than the average individual's (one-tailed z-test, p<0.001), 50% smaller than the survey mean, and 70% smaller than GPT-4 (p. 5, §V–VI). 3. The CSI estimate is a sentiment-weighted mean over the 10 options computed by NLP and "not shown to participants in real-time" (p. 4, §IV). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: LLM agents link small-group deliberations at scale | (c) AI and group ideation/decisions: AI-linked group deliberation gave a more accurate group judgement on one estimation item - **Caveats:** Vendor-authored pilot; a single estimation item; the CSI vs survey-mean and CSI vs GPT-4 comparisons are single numbers with no significance test; the order and allocation of conditions for the same 241 people are not detailed in the text read; the CSI estimate depends on the system's own weighting scheme. - **Screening check:** PASSES (empirical; humans in groups interacting with LLM agents; method and n reported); decision-quality outcome. #### OA2213 — Conversational Swarm Intelligence, a pilot study (~25 students) - **Citation:** Rosenberg, L., Willcox, G., Schumann, H., Bader, M., Mani, G., Sagae, K., Acharya, D., Zheng, Y., Kim, A., & Deng, J. (2023). Conversational Swarm Intelligence, a Pilot Study. arXiv:2309.03220. https://doi.org/10.48550/arxiv.2309.03220 - **Type:** preprint — peer-reviewed: no (preprint) - **Full text read:** yes (https://arxiv.org/pdf/2309.03220, 4 pages) - **Method:** Within-group pilot: 10 questions put to the same group, half in a standard chat room and half in CSI; contribution counts compared (p. 3, Pilot Study). - **n:** One group of "roughly 25" users (student volunteers from a Carnegie Mellon summer class, per Acknowledgements); CSI trials used 5 subgroups of ~5 (p. 3). - **Setting:** Text chat. Tasks: 4 chess-puzzle best-move questions and 1 open-ended AI-ethics question per condition, ≤5 minutes each. GPT-3.5 agents observed one subgroup and passed summaries to another at one-minute intervals (p. 3). - **Findings:** 1. Contributions (suggestions, explanations, agreements, disagreements, questions) were 30% higher with CSI than in standard chat (mean 9.48 vs 7.28 per user; p<0.05) (p. 3, Table 1). 2. Variance in contribution was 7.2% lower in CSI (87.8 vs 94.6), which the authors read as more even participation (p. 3, Table 1). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: LLM relays between subgroups as a participation structure | (c) AI and group ideation/decisions: more contributions; answer quality not reported - **Caveats:** Very small pilot, student volunteers, 4-page preprint; no accuracy outcomes for the chess tasks; statistical test not specified; vendor co-authors. - **Screening check:** PASSES (empirical; group with LLM agents; method and n reported) — pilot quality. #### OA2540 — Large-scale group brainstorming: CSI vs traditional chat (~75-person groups) - **Citation:** Rosenberg, L., Schumann, H., Dishop, C. R., Willcox, G., Woolley, A. W., & Mani, G. (2024). Large-scale Group Brainstorming using Conversational Swarm Intelligence (CSI) versus Traditional Chat. arXiv:2412.14205. https://doi.org/10.48550/arxiv.2412.14205 - **Type:** preprint — peer-reviewed: no (preprint) - **Full text read:** yes (https://arxiv.org/pdf/2412.14205, 6 pages) - **Method:** Counterbalanced within-group comparison: each of 2 groups did one Alternative Use Task (AUT) in a single large chat room and another in CSI, in reversed order, followed by a 7-item comparative preference survey; one-proportion z-tests with Bonferroni adjustment (p<0.0014) (p. 4–5, §2–4). - **n:** 2 groups of about 75 participants from a commercial sample provider; 147 surveys collected. In CSI, 15 subgroups of 5 people, each with one AI agent (p. 4–5). - **Setting:** Text-based Thinkscape vs a Slack/Discord-like single chat room. AUT tasks: alternative product uses for surplus traffic cones or toilet plungers, 12 minutes each. The LLM "conversational surrogates" only relayed ideas between subgroups and did not add AI-generated ideas (p. 4). - **Findings:** 1. A significant majority preferred CSI on all 7 questions (each p<0.0014 after Bonferroni), with support "ranging between 66% and 88%" and 75% preferring CSI overall (p. 5, §3–4). 2. Participants rated CSI more productive, more collaborative and better at surfacing quality answers, and "over 80%" reported feeling more heard and more ownership and buy-in (p. 5, §5). 3. No objective measure of idea quantity or quality is reported; outcomes are subjective preferences only (p. 4–5). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: LLM agents structure large-group brainstorming | (c) AI and group ideation/decisions: participants preferred AI-linked brainstorming, but ideation output was not measured - **Caveats:** Vendor-authored preprint; paid panel; outcomes are self-report comparisons only; per-question percentages appear only in Fig. 4 (not extracted as text); tasks were confounded with condition within each group (counterbalanced across groups). - **Screening check:** PASSES (empirical; large human groups brainstorming with LLM agents; method and n reported). #### OA1308 — Conversational Swarm Intelligence (Thinkscape) for 75-person brainstorming - **Citation:** Rosenberg, L., Schumann, H., Dishop, C., Willcox, G., Woolley, A., & Mani, G. (2025). Large-Scale Group Brainstorming and Deliberation Using Swarm Intelligence and Generative AI. Proceedings of the 27th International Conference on Enterprise Information Systems (ICEIS 2025), Vol. 1, pp. 867–872. SciTePress. https://doi.org/10.5220/0013379800003929 - **Type:** conference (short paper) — peer-reviewed: unclear (ICEIS proceedings; review not described in text) - **Full text read:** yes (https://doi.org/10.5220/0013379800003929; PDF from SciTePress) - **Method:** Two-condition within-subjects comparison with counterbalanced order: each group of ~75 did two 12-minute Alternative Use Task brainstorms (traffic cones, toilet plungers) — once in a single large text chat room, once in the CSI platform Thinkscape (≈15 subgroups of 5 people, each with one LLM "conversational surrogate" agent relaying ideas between subgroups every 30–60 s); post-session forced-choice preference survey (7 questions); one-proportion z-tests with Bonferroni adjustment (p < 0.0014) (§2–3, p. 870–871). - **n:** Two groups of approximately 75 participants each (sourced from a commercial sample provider); 147 surveys collected; 30-minute sessions (§2, §3, p. 870–871). - **Setting:** Online text-based platform (Thinkscape by Unanimous AI) vs standard chat room; AI agents did not introduce their own ideas, only passed ideas/opinions between subgroups (§2, p. 870). LLM model not stated (not verified). - **Findings:** 1. A significant majority preferred Thinkscape on all 7 survey questions (productive, heard, collaborative, surfaced better answers, buy-in, ownership, overall), with support "ranging between 66% and 88%", and 75% preferring Thinkscape overall; all significant at the Bonferroni-adjusted level (§3–4, p. 871, Fig. 4). 2. Authors state "over 80%" of participants reported feeling "more heard" and "more ownership" and "more buy-in" with CSI (§5, p. 871) — per-question figure values not given in text (not verified beyond this statement). 3. The paper cites the authors' earlier CSI work (not this study's data): 51% more content (p<0.001) and 37% less contribution inequality vs centralized chat in 48-person groups (§1, p. 869) — secondary, not a finding of this study. - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: LLM agents act as automated connective facilitation layer enabling 75-person real-time brainstorming/prioritization | (c) AI and group ideation/decisions: only subjective preference measured; no measures of idea quantity/quality or decision quality in this study. - **Caveats:** Outcomes are self-reported comparative preferences only; no objective ideation metrics; authors include Unanimous AI employees (developer of Thinkscape) — conflict of interest; commercial panel sample; LLM model unspecified; short paper. - **Screening check:** Meets criteria (empirical; large human groups interacting in real time with LLM agents; method and n reported; English). #### C23 — Conversational Swarm Intelligence amplifies accuracy of networked group deliberations - **Citation:** Rosenberg, L., Willcox, G., Schumann, H., & Mani, G. (2024). Conversational Swarm Intelligence amplifies the accuracy of networked groupwise deliberations. In 2024 IEEE 14th Annual Computing and Communication Workshop and Conference (CCWC), pp. 0086–0091. https://doi.org/10.1109/CCWC60891.2024.10427807 - **Type:** conference (short paper) — peer-reviewed: yes (IEEE conference; read as author preprint) - **Full text read:** yes (arXiv preprint, https://arxiv.org/pdf/2401.04112v1) - **Method:** Pilot within-group comparison over 11 weekly sessions: individuals first complete a survey (individual choice; Wisdom-of-Crowd aggregate), then the same group deliberates in real-time text chat on the Thinkscape CSI platform; paired t-test by session, bootstrap, pairwise comparison of 55 selections (§IV–V, p. 3–4). - **n:** 11 sessions with groups of 25–30 participants each week (Amazon Mechanical Turk, self-identified NFL fans); exact total unique participants not reported (§IV, p. 3). - **Setting:** Selecting 5 Fantasy Football players under a fixed budget ($32,500), 4 minutes per position (~20 min session). CSI splits a large group into small chat subgroups, each with an LLM-powered conversational agent that distils the local arguments and relays neighbouring rooms' content as first-person dialogue (§III, p. 2). Specific LLM not named in the paper (not verified). - **Findings:** 1. The CSI group averaged 86.8 points per session vs. 77.2 for the median individual (paired t-test p = 0.020) and 74.2 for the Wisdom-of-Crowd survey roster (p < 0.001); CSI beat the median individual in 8 of 11 sessions (§V, Table 1, p. 4). 2. On average CSI outperformed 66% of individual rosters (bootstrap 95% CI 61.4%–69.7%) (§V, p. 4). 3. Of 55 player selections, CSI and WoC agreed on 40 (73%); where they differed, CSI's pick outperformed 13 of 15 times vs. 2 for WoC (p = 0.004) (§V, Table 2, p. 4). - **Relevance:** (a) — | (b) LLM agents as cross-subgroup relays enabling large-scale real-time deliberation (a facilitation/structuring function) | (c) group decision accuracy with AI-mediated deliberation exceeded median individual and survey aggregation - **Caveats:** Pilot study by the platform vendor (Unanimous AI); MTurk sample; no chat-without-AI control in this study (comparison is against individual surveys); participant totals and LLM unspecified; prior-study figures cited in §III are not from this study. - **Screening check:** Meets criteria — empirical, groups of 25–30 humans deliberating with LLM-agent mediation, method and n (sessions/group size) reported. #### C84 — Conversational Swarm Intelligence with LLM "Infobots" for group decisions - **Citation:** Rosenberg, L. B., Schumann, H., Dishop, C. R., Willcox, G., Woolley, A. W., & Mani, G. (2024). Conversational Swarms of Humans and AI Agents enable Hybrid Collaborative Decision-making. 2024 IEEE 15th Annual Ubiquitous Computing, Electronics & Mobile Communication Conference (UEMCON), pp. 257–263. https://doi.org/10.1109/uemcon62879.2024.10754763 (arXiv:2410.03690) - **Type:** conference — peer-reviewed: yes (IEEE conference). The full text read is the arXiv version. - **Full text read:** yes (https://arxiv.org/pdf/2410.03690) - **Method:** Field study over 5 weeks with 2 sessions per week (10 sessions). Each week one session used standard CSI and one used CSI plus an LLM Infobot (OpenAI Assistants API). Group rosters were compared with individual survey rosters and a Wisdom-of-Crowd aggregate (paired t-test, Wilcoxon, bootstrap), followed by an exit survey (p. 3–6 of PDF, §III–V). - **n:** 10 sessions with groups of about 25 participants each (the exact total is not reported). Each group was split into about 5 subgroups of about 5, linked by LLM "Surrogate Agents". 5 sessions had Infobots (p. 3–4 of PDF). - **Setting:** Self-identified baseball fans, online, on the Thinkscape CSI platform (Unanimous AI). The task was choosing 5 players for a real DraftKings fantasy baseball contest within a budget, with 5 min 30 s per position and sessions of about 30 min. The Infobot answered @infobot factual queries visible to the subgroup (p. 3–4 of PDF). - **Findings:** 1. CSI group rosters scored 62.4 points vs. 47.3 for the median individual (p = 0.004) and 43.7 for the WoC aggregate (p = 0.008). They beat 73% of individual rosters vs. 39% for WoC (p = 0.011) (p. 4 of PDF, Table 1). 2. Infobot sessions reached the 74th vs. 72nd percentile, which was not significant. Per-participant variance in contribution was lower with the Infobot (12.2 vs. 15.0 characters/min, p = 0.039), and the authors read the Infobot as an "equalizer". Subgroups averaged 4.1 Infobot queries per player selected (p. 4–6 of PDF, Table 2, §5.1). 3. 86.5% agreed "Our decisions were stronger because of information provided by the Infobot" (the abstract says 85%). Over 90% agreed with positive statements about CSI, e.g. 97.3% "able to share my views openly without being judged" (p. 6 of PDF, Tables 4, 6). - **Relevance:** (a) — | (b) AI agents relay content between subgroups and act as an in-discussion information resource in large-group deliberation | (c) group decisions beat individuals and WoC. The LLM info agent balanced participation but gave no significant accuracy gain. - **Caveats:** Authors from Unanimous AI (vendor). Only 10 sessions. The exact number of participants is not given. Sports-forecasting task. Short conference paper. The abstract (85%) and table (86.5%) disagree. Overlaps with C85 (same group and platform). - **Screening check:** Meets criteria (empirical, groups of about 25 humans deliberating with LLM agents, method and n (sessions) reported, English). #### C85 — Hyperchat AI conversational forecasting (MLB) - **Citation:** Schumann, H., Rosenberg, L. B., Mani, G., & Willcox, G. (2025). Conversational Collective Intelligence (CCI) using Hyperchat AI in a Real-world Forecasting Task. 2025 IEEE CHIUXiD conference, pp. 95–100. https://doi.org/10.1109/chiuxid68326.2025.11323746 (arXiv:2511.03732) - **Type:** conference — peer-reviewed: yes (IEEE conference). The full text read is the arXiv version. - **Full text read:** yes (https://arxiv.org/pdf/2511.03732) - **Method:** Field forecasting study with no non-AI control condition. Group forecasts were benchmarked against Vegas betting odds. An LLM scored each participant's view and conviction in real time, and games were classified as High or Low Confidence (p. 2–5 of PDF, §II–III). - **n:** 15 sessions of 20–25 Prolific participants each (about 24 per the abstract; the exact total is not reported). 60 games were forecast and 59 analysed (one rained out) (p. 2 of PDF). - **Setting:** Self-identified baseball fans (paid about $6.50 per session) on Thinkscape.ai. Each group was split into 5 subgroups of 4–5 people, each with an LLM "Conversational Surrogate" agent that monitors its subgroup, extracts insights and relays them to the other subgroups. Deliberation lasted 5 minutes per game in sessions of about 30 min (p. 2–3 of PDF). - **Findings:** 1. High Confidence picks (27 of 59 games, 46%) were 78% accurate vs. 57% Vegas implied odds (p = 0.020), 95% CI 59.2–89.4%. Low Confidence picks had a 95% CI of 25.5–57.7% (p. 3 of PDF). 2. High Confidence picks reached through above-average conversation rates (16 games) were 88% accurate vs. 57% Vegas odds (p = 0.010), which the authors read as showing that interactive deliberation matters (p. 4–5 of PDF). 3. Betting against the spread on High Confidence picks would have given 63% accuracy and a 46% ROI (p = 0.037) (p. 4 of PDF). - **Relevance:** (a) — | (b) LLM agents relay insights between subgroups so that a large group can deliberate in real time | (c) AI-networked group deliberation produced forecasts more accurate than betting markets on high-confidence items - **Caveats:** Vendor authors (Unanimous AI). No control group without AI, so the AI contribution is not isolated. Sports task. Short paper. The abstract says 8 weeks, the body "seven consecutive weeks". Overlaps with C84. - **Screening check:** Meets criteria (empirical, groups of 20–25 humans deliberating with LLM agents, method and n reported, English). #### C672 — Hyperchat AI: surrogate-agent swarm for large-group conversational forecasting - **Citation:** Rosenberg, L., Schumann, H., Mani, G., & Willcox, G. (2026). Conversational Forecasting Across Large Human Groups Using a Swarm of Surrogate AI Agents. 2026 2nd International Conference on Human-AI Interaction and Experience Design (HAXD), Valencia, pp. 101–107. IEEE. https://doi.org/10.1109/haxd70072.2026.11620928 (read as arXiv:2604.09570v2) - **Type:** conference — peer-reviewed: unclear (IEEE conference; text read is the authors' preprint) - **Full text read:** yes (https://arxiv.org/pdf/2604.09570, preprint; page numbers below = PDF pages) - **Method:** Field forecasting study without a concurrent control group; group forecasts from AI-mediated text deliberation compared against Vegas odds (binomial test), participants' own pre-deliberation individual forecasts (bootstrap), and Polymarket (§III–IV). - **n:** 56 NBA games forecast over 12 weeks (50 scorable; 6 toss-ups); groups of 25–30 Prolific participants per session, split into 5–6 subgroups of ~4–5, four games per ~20-minute session; 43 participants took part in ≥12 game forecasts (pp. 2–5, §III–IV). Total unique participants not stated (not verified). - **Setting:** Self-identified basketball fans (Prolific, ~$6.50 per session) using Unanimous AI's Thinkscape platform; one LLM-based "Conversational Surrogate" agent per subgroup passes insights between subgroups (selected by a "Deliberative Matching Engine" to maximally challenge the receiving subgroup); an LLM scores each participant's stance and conviction from their dialogue; human moderator posts the prompt; specific LLM not named (pp. 2–4, §III). - **Findings:** 1. AI-mediated groups were correct on 31 of 50 forecasts (62.0%) vs. 50% Vegas benchmark (p = 0.059), implied ROI +18.4% (p. 4–5, §IV, Table I). 2. Group accuracy (62%) exceeded the mean individual pre-deliberation accuracy of 52% and outperformed 36 of 43 regular participants (84%); vs. bootstrapped individual forecasts p = 0.071 (p. 5, §IV). 3. Accuracy tracked conversation rate: forecasts above the 25th percentile of characters/minute reached 68.4% on 38 games (p = 0.017 vs. Vegas; vs. Polymarket 54.8%, p = 0.062), while the lowest-quartile forecasts had 41.7% on 12 games (pp. 5–6, §IV–V). - **Relevance:** (a) — | (b) LLM agents run/connect parallel subgroup deliberations at scale (AI-mediated large-group deliberation) | (c) AI-mediated deliberation produced group forecasts more accurate than members' individual forecasts, though mostly at marginal significance - **Caveats:** Authors are affiliated with the vendor (Unanimous AI); no control condition with unmediated deliberation or plain aggregation of the same participants; headline p-values mostly > .05; subgroup filtering by conversation rate is post hoc; LLM not named; preprint text. - **Screening check:** Meets criteria — empirical field study, groups of 25–30 humans deliberating together via LLM agents that relay content between subgroups; method and n reported; English. #### R86 — Conversational Swarm Intelligence and group IQ - **Citation:** Rosenberg, L. B., Willcox, G., Schumann, H., & Mani, G. (2024). Towards Collective Superintelligence: Amplifying Group IQ Using Conversational Swarms. Proceedings of the 16th International Conference on Agents and Artificial Intelligence (ICAART 2024), SCITEPRESS, pp. 759-766. https://doi.org/10.5220/0012687500003690 - **Type:** conference — peer-reviewed: yes - **Full text read:** yes (arXiv version, https://arxiv.org/pdf/2401.15109; pages = arXiv PDF pages) - **Method:** comparison of real-time networked group deliberation on the Thinkscape CSI platform vs a baseline of individuals answering by survey; Raven's Advanced Progressive Matrices items; paired t-test on question-by-question accuracy (§II-IV, p. 3-4) - **n:** baseline group of 35 individuals; CSI groups of approximately 35 people each, split into 7 subgroups of 5, each subgroup with an LLM-powered agent (p. 3). The number of CSI sessions/groups is not stated (not verified). - **Setting:** paid participants from a commercial sample provider; text chat deliberation, up to 4 minutes per question; LLM-powered agents pass insights between subgroups and do not introduce their own content (p. 1-3) - **Findings:** 1. CSI groups averaged 80.5% correct vs 45.7% for the average individual, corresponding to an "effective IQ" of 128 vs 100 (p < 0.001) (Abstract; §IV, p. 4). 2. CSI groups also outperformed a statistical aggregation of individual answers (IQ 115, p < 0.01), and none of the 35 baseline participants performed as well as the CSI group (§V, p. 4). 3. The authors report that the LLM agents only relayed conversational content across subgroups, so the gain is attributed to the networked deliberation structure (p. 3). - **Relevance:** (a) — | (b) LLM agents structure/connect large-group real-time deliberation | (c) AI-mediated group deliberation improved group decision accuracy on IQ items - **Caveats:** authors are employees of the platform vendor (Unanimous AI); number of sessions not reported; baseline individuals vs groups (no human-only group deliberation control); arXiv abstract says the baseline averaged 45.6% while the text says 45.7%. - **Screening check:** meets criteria — empirical comparison, multi-human groups deliberating with LLM involvement, method and n (approximate) stated, English. Borderline on sample reporting (session count missing). #### C659 — StoryBuilder: human-AI narrative synthesis of civic engagement input - **Citation:** Overney, C., Jiang, H., Haider, U., Moe, C., Mangat, J., Pantano, F., McMillian, E. G., Riggins, P., & Gillani, N. (2026). Human-AI Narrative Synthesis to Foster Shared Understanding in Civic Decision-Making. Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26), pp. 1–24. https://doi.org/10.1145/3772318.3791570 (read as arXiv:2509.19643v1) - **Type:** conference — peer-reviewed: yes (CHI); text read is the arXiv preprint version, may differ from the published version - **Full text read:** yes (https://arxiv.org/pdf/2509.19643, v1, 23 Sep 2025) - **Method:** System building plus mixed-methods evaluation: human technical evaluation of 20 generated stories; four-month field deployment with usage analytics; remote user-study interviews (inductive thematic analysis); preregistered between-subjects online experiment (4 narrative-composition conditions incl. control; one-way ANOVAs with Tukey tests) (§3.2.6, §4.1–4.3). - **n:** Input: 2,480 community feedback segments (56% survey responses, rest listening-session transcripts) → 190 AI-generated stories, 124 deployed (66 dropped, 44 edited, 80 unchanged) (p. 6 §3.2.1; p. 7 §3.2.5). Technical evaluation: 20 stories, 3 annotators each (six authors) (p. 7–8 §3.2.6). Field: 2,183 sessions (p. 11 §4.1.4). User study: 21 community members (p. 11 §4.2.1). Experiment: 200 Prolific parents recruited, 198 analysed (pp. 18–19 §4.3). - **Setting:** Contentious school rezoning process in a US district of ~50,000 students; pipeline using GPT-4o-mini, GPT-4o and Claude 3.5 Sonnet (DSPy) to turn community input into first-person composite stories with citations, reviewed by humans and shown in the mobile StorySharer interface (pp. 3, 6–7, §3.2). - **Findings:** 1. Human rating of 20 LLM stories: highly readable (0.99 ± 0.09), believable (0.97 ± 0.18), relevant (0.88 ± 0.33), coherent (0.84 ± 0.37), citations mostly appropriate (0.92 ± 0.28), but citation attribution accuracy only 0.55 ± 0.50 — authors describe "citation hallucination" (p. 8, §3.2.6). In human review, 66 of 190 stories were dropped (p. 7, §3.2.5). 2. Preregistered experiment: experience-heavy (scene-dominant) stories elicited more respect (F(3,194) = 5.02, p = .002, η² = .072; M = 3.72 vs 3.29 theme-dominant) and trust (F(3,194) = 3.02, p = .031, η² = .045) than opinion-heavy stories; no condition effect on stance change or focus of consideration (all p > .12) (p. 19, §4.3.4). 3. Field deployment: 51% mobile sessions, mean 6.23 min (SD 59.98), but only 233 feedback responses from 56 sessions (3.3% of active users); interview participants largely accepted AI use with human review (only 5 of 21 gave measured critiques), whereas several online-experiment participants reacted with aversion after learning stories were AI-written (pp. 11, 14, 19, §4.1.4, §4.2, §4.3.4). - **Relevance:** (a) — | (b) LLM performing synthesis of community-engagement input (incl. listening-session transcripts) to be shared back to the community; human-in-the-loop quality issues (citation errors) | (c) framing of AI synthesis affects respect/trust but not stances - **Caveats:** Readers engaged individually with stories — no group interaction with the AI; experiment on Prolific with parents from other districts; field analytics anonymous; arXiv v1 read, not the CHI camera-ready. - **Screening check:** Borderline. Empirical, method and n reported, English, LLM involved. It does not involve ≥2 humans interacting as a group with the AI, and respondents are not facilitators; it passes only if LLM synthesis of civic engagement input (partly listening-session transcripts) is counted as "an LLM performing a facilitation function on group-discussion data". Under a strict reading: FAILS: EX-INDIV/EX-TOPIC (individual readers of AI-synthesised civic feedback; no facilitated group session). ## 6. Education settings with AI in structured group work, and reviews #### C283 — Team-LLM collaboration in AI-infused classroom debate - **Citation:** Zhang, Z., Sun, B., & An, P. (2025). Breaking Barriers or Building Dependency? Exploring Team-LLM Collaboration in AI-infused Classroom Debate. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI '25), Yokohama, pp. 1–19. https://doi.org/10.1145/3706598.3713853 (arXiv:2501.09165) - **Type:** conference — peer-reviewed: yes. The full text read is the arXiv version. - **Full text read:** yes (https://arxiv.org/pdf/2501.09165) - **Method:** Qualitative field study in a university course: classroom observation, recordings, logs of team–ChatGPT dialogue, and individual semi-structured interviews. Thematic analysis by two coders plus a reviewer produced 10 themes and 29 sub-themes (p. 4–6 of PDF, §3). - **n:** 22 students (11 female, 11 male). 3 rounds of 5-vs-5 debates over 4 weeks (about 1.5 h each), with non-debating students as the audience. Interviews averaged 30.75 min, 1,022.15 min of audio in total. Teams sent 59 messages to ChatGPT (8,181 words) and received 37,752 words in reply (p. 4–6 of PDF). - **Setting:** Design History course at SUSTech (China). Each team could use ChatGPT 3.5 on a single shared laptop during preparation and cross-examination (p. 4 of PDF, §3.1). - **Findings:** 1. Teams developed divisions of labour around the AI (e.g. one member prompts ChatGPT while another summarises and extracts). Some groups set fixed roles and others flexible ones (p. 8 of PDF, §4). 2. Advantages: the AI reduced social anxiety about asking questions, broke communication barriers, and scaffolded novices (p. 9–10 of PDF, §5.1–5.2). 3. Risks: information overload, since ChatGPT responses averaged 640 words (SD = 314) against 139-word prompts, too many ideas and irrelevant content under time pressure. Cognitive dependency grew over time ("AI gives so many answers at once that our team skips the thinking phase") and attention shifted from listening to opponents towards the AI output (p. 6, p. 10–11 of PDF, §5.3–5.4). - **Relevance:** (a) — | (b) LLM as an in-session team resource during a structured, time-boxed group format (debate) | (c) the AI lowers participation barriers but risks overload and dependency that reduce independent team thinking - **Caveats:** Single course, n = 22 students, qualitative with no comparison condition. ChatGPT 3.5 (older model), limited to one device per team. - **Screening check:** Meets criteria (empirical, teams of 5 humans interacting with an LLM during group sessions, method and n reported, English). #### C334 — Teachers' perspectives on a voice LLM peer agent in group work (Phoenix) - **Citation:** Ravi, P., Stevens, C., Flamia Azevedo, B., David, J., Hanks, B., Abelson, H., Lin, G. C., & Anderson, E. (2026). Exploring Teachers' Perspectives on Using Conversational AI Agents for Group Collaboration. In Artificial Intelligence in Education (AIED 2026), Lecture Notes in Computer Science, pp. 621–630. https://doi.org/10.1007/978-3-032-29763-1_43 (preprint arXiv:2602.07142) - **Type:** conference (LNCS) — peer-reviewed: yes - **Full text read:** yes (arXiv v2 accepted version, https://arxiv.org/pdf/2602.07142; page numbers refer to that PDF) - **Method:** Exploratory qualitative study within a 2.5-hour professional-development workshop: teachers in groups of three played three activities with the agent (10-min icebreaker, 20-min consensus-building "sinking ship" ranking, 15-min open-ended brainstorming), followed by an open-ended reflection survey or focus groups; inductive thematic analysis with social moderation (§4, pp. 5–6). - **n:** 33 K-12 STEM teachers in 11 groups of three; 22 completed the reflection survey, 11 took part in 3 focus groups of 3–4 (§4.1–4.2, pp. 5–6). - **Setting:** Teacher PD programme (26 US, 7 international teachers). "Phoenix": voice-based, non-embodied peer agent on a laptop at the table; GPT-4.1-mini as core model, Google speech-to-text, Azure text-to-speech, GPT-4.1-mini gatekeeper deciding when Phoenix should respond; prompted to speak ~20 words, build on group ideas, ask clarifying questions (§3, pp. 3–4). - **Findings:** 1. Perceived value was mixed: teachers valued Phoenix surfacing overlooked considerations, speaking up "when others may not" and recapping previous decisions (memory aid), but others said it "only summarized things we already said", was overly agreeable or adamant, and lacked situational awareness (§5.1, p. 7). 2. Roles shifted between peer/thought partner, efficiency tool/reference material, and "source of truth and arbitration" — groups asked it to act as a neutral tie-breaker during disagreement ("it changed how we prioritized") and it helped break deadlocks; trust depended on response timing (slow responses) and stakes (§5.1, pp. 7–8). 3. Several groups excluded the agent once its ideas did not exceed theirs ("Sometimes we whispered so it wouldn't hear us"); teachers envisioned classroom roles as TA/tutor, group monitor, "sparring partner", but worried about interruptions of productive dialogue, overreliance, privacy and threats to teacher authority (§5.1–5.2, pp. 9–10). - **Relevance:** (a) facilitator use of AI: teachers (group-work facilitators) as respondents on AI in collaboration | (b) AI planning/running/supporting group sessions: LLM voice agent participating in live small-group consensus and brainstorming tasks; turn-taking/interruption issues | (c) AI and group ideation/decisions: qualitative reports of agent as tie-breaker shifting group rankings; no measured outcomes - **Caveats:** Short exposure (~45 min), qualitative only, no outcome measures; single workshop; researcher-facilitated groups; short 10-page paper. - **Screening check:** Meets criteria — empirical qualitative study, groups of 3 humans interacting with an LLM agent in structured group tasks (consensus-building, brainstorming), method and n reported, English. (Educational context, but not individual classroom use; group process explored qualitatively only.) #### C584 — Masked GenAI chatbot in PyramidApp collaborative chat rooms - **Citation:** Gutiérrez-Ferré, A., Hernández-Leo, D., & Sánchez-Reina, J. R. (2024). Generative AI Chatbot in PyramidApp: Students' Behaviors and Design Principles. In Collaboration Technologies and Social Computing (CollabTech 2024), Lecture Notes in Computer Science 14890, pp. 248–256. Springer. https://doi.org/10.1007/978-3-031-67998-8_19 - **Type:** conference (LNCS, work in progress) — peer-reviewed: yes - **Full text read:** yes (author manuscript in UPF e-Repositori, http://hdl.handle.net/10230/68391; page numbers = manuscript pages 1–8, not the LNCS pagination) - **Method:** Quasi-experimental within-subjects study embedded in 9 online PyramidApp activities in two engineering courses; a "masked" chatbot joined some groups as a pseudonymous participant via Wizard-of-Oz (research assistant relaying between PyramidApp and ChatGPT web interface / custom GPTs) in individual submission, peer rating and chat-room improvement phases; log analysis comparing chatbot-assisted and non-assisted rooms with descriptive and inferential statistics (§2–3, pp. 2–4). - **n:** 291 students; 9 activities; 105 chat rooms of 4–6 students; 284 individual submissions; rating comparison restricted to 27 students in groups with the chatbot; chat statistics e.g. 1,253 messages in non-assisted vs. 141 in assisted rooms (§2, p. 2; §3.1, p. 4; Table 4, p. 5). - **Setting:** Ordinary engineering class sessions at Universitat Pompeu Fabra (Barcelona); case-study tasks (coding, non-functional requirements, ethics); students unaware the pseudonymous member was AI; ChatGPT (model version not verified) (§2.1–2.2, pp. 2–3). - **Findings:** 1. Chatbot rated peers' answers lower than students did (M = 3.48, SD 1.02 vs. 4.22, SD 0.97, p < 0.001) but its ratings were positively correlated with students' ratings (slope t-test p = 0.025) (§4.1, p. 5). 2. Chatbot's own submissions were rated higher than students' (M = 4.11 ± 0.84 vs. 3.91 ± 1.06) but not significantly (p = 0.330) (§4.2, p. 5; Table 3). 3. In chatbot-assisted rooms, students wrote fewer short (1–2 word) messages per member (2.53 vs. 1.41, p = 0.003), fewer messages per member (5.83 vs. 3.90, p = 0.011) and longer messages (3.78 vs. 4.58 words, p = 0.031); short messages fell from 46% to 36%; total word count per member was lower (23.30 vs. 16.43, p = 0.064) (§4.3, p. 5; Table 4). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: LLM (WoZ-relayed) acting as a participant/feedback giver in scripted small-group discussion; design principles for rating and improvement phases (pp. 6–7) | (c) AI and group ideation/decisions: changes in group discussion behaviour (fewer, longer messages); no measure of decision quality or learning - **Caveats:** Work-in-progress paper; Wizard-of-Oz relay with occasional manual reformatting; students deceived about AI presence; very unbalanced assisted vs. non-assisted samples (e.g., 13 vs. 60 rooms in room-level metrics); message length as quality proxy; classroom setting. - **Screening check:** Meets criteria (borderline EX-EDU) — ChatGPT-generated contributions entered live group discussions of 4–6 students in a structured collaboration script and group-level process was measured; method and n reported; English. #### C585 — ChatGPT as "team member" in collaborative lesson design (crossover) - **Citation:** Agostini, D., Serbati, A., Picasso, F., & Lipnevich, A. A. (2026). When ChatGPT joins the team: a mixed-methods study of AI-mediated collaborative lesson design. Frontiers in Education, 11, 1751618. https://doi.org/10.3389/feduc.2026.1751618 - **Type:** journal — peer-reviewed: yes - **Full text read:** yes (https://public-pages-files-2025.frontiersin.org/journals/education/articles/10.3389/feduc.2026.1751618/pdf; CC BY) - **Method:** Mixed-methods quasi-experimental counterbalanced crossover within one 240-min session: groups randomly assigned to use ChatGPT in Task 1 (GT1) or Task 2 (GT2), instructed to "treat and address ChatGPT as if it were a group member"; individual 13-item knowledge test at T0/T1/T2 (Wilcoxon, rank-biserial r); group artifacts scored by two expert raters on a 3-criterion rubric (ICC(2,1) = 0.934 / 0.944); inductive thematic analysis of written group reflections in Atlas.ti (§4, pp. 5–8). - **n:** 102 post-graduate special-education teacher trainees (82 women; M age 38.87, SD 8.01) in 21 groups of 4–6; 82 students completed all three knowledge tests (38 GT1, 44 GT2) (§4.1, p. 5; §5.1, p. 9). - **Setting:** Italian 1-year specialisation programme for special-education teachers, course on instructional design with technology for inclusion; Task 1 = select and justify an inclusive technology for a learning objective; Task 2 = full inclusive lesson design; ChatGPT 3.5 (§4.2–4.3, pp. 5–7). - **Findings:** 1. Knowledge test: GT1 improved after the AI-assisted task (overall p = 0.015, r_rb = −0.429; technology items p = 0.001, r_rb = −0.692) and stayed stable after AI withdrawal; GT2 improved after its AI task (overall p = 0.023, r_rb = −0.371; technology p = 0.022, r_rb = −0.442); Treatment × Sequence interaction not significant (all ps ≥ 0.387) (§5.1.1–5.1.3, p. 9). 2. Group artifact quality rose from Task 1 to Task 2 (mixed ANOVA main effect of Task F(1,19) = 6.33, p = 0.021; no sequence effect, F = 0.27, p = 0.612; no interaction, p = 0.776); final Task 2 quality did not differ between early-AI (M = 14.75, SD 2.92) and late-AI groups (M = 14.38, SD 4.43), p = 0.82, d = 0.09 (§5.1.4, p. 9). 3. Qualitative: 43% of groups used ChatGPT as an "oracle"/expert guide at least once; in Task 2 all AI-using groups reported engagement with the design process with AI (100%) and 75% reflected on AI integration; weaknesses included "less motivated to introduce personal creativity, adopting more standard solutions" and "less cooperation among group members and mental commitment" (pp. 13–14; Table 10, p. 15); abstract reports reduced peer-to-peer idea-building in some groups. - **Relevance:** (a) facilitator use of AI: — (participants are trainee teachers designing lessons, not facilitating) | (b) AI planning/running/supporting group sessions: ChatGPT positioned as a peer "team member" in face-to-face group design work (planning of learning activities) | (c) AI and group ideation/decisions: no detectable quality loss in group products after AI withdrawal; qualitative signs of standardisation and reduced peer idea-building - **Caveats:** Single session; quasi-experimental with only 21 groups; knowledge test has very low internal consistency (α 0.278–0.387); ChatGPT 3.5; group-level quality differences small; educational (teacher-training) sample; group split per sequence not fully verified. - **Screening check:** Meets criteria (borderline EX-EDU) — groups of 4–6 humans collaborating with ChatGPT as a designated group member in a structured group design task, with group-level artifact quality and group process measured; method and n reported; English. #### R1194 — GenAI chatbot in PyramidApp CSCL group activities - **Citation:** Sánchez-Reina, J. R., Gutiérrez-Ferré, A., & Hernández-Leo, D. (2025). Measuring the Effect of GenAI Chatbots in CSCL environments: A Study with PyramidApp. Proceedings of the International Conference on Computer-Supported Collaborative Learning (CSCL 2025), 622-624. ISLS. https://doi.org/10.22318/cscl2025.114109 - **Type:** conference (short paper) — peer-reviewed: yes (ISLS proceedings; not verified further) - **Full text read:** yes (https://repository.isls.org//bitstream/1/11935/1/CSCL2025_622-624.pdf) - **Method:** quasi-experimental within-subject study across nine collaborative online activities; chatbot simulated via Wizard of Oz with responses generated by ChatGPT (templates, later custom GPTs); rubric coding of knowledge building and contribution quality; chat analysis at message, individual and group level (Method, p. 622) - **n:** 284 individual and 105 group submissions analysed across 9 activities (p. 623); number of students not reported - **Setting:** students in PyramidApp (pyramid collaborative learning flow: individual reflection -> rating -> group discussion/improvement in live chat); masked ChatGPT-based "peer" contributed an individual submission, ratings and live-chat comments (Table 1, p. 622) - **Findings:** 1. 52.9% (N = 9) of groups incorporated ChatGPT's contributions into their final submissions; of these 55.6% used some content, 22.2% copied with added explanations, 22.2% copied exactly (Results, p. 623). 2. Group submissions: precision 2.11 vs 2.31 (p = 0.254) and confusion 0.56 vs 0.44 (p = 0.516), groups without vs with ChatGPT — not significant (Table 2, p. 623). 3. In chat rooms with the chatbot, members sent fewer messages (5.83 vs 3.90 per member, p = 0.011) but longer ones (3.78 vs 4.58 words, p = 0.031) and fewer short 1-2-word messages (2.53 vs 1.41, p = 0.003) (Table 3, p. 623). - **Relevance:** (a) — | (b) an LLM-generated participant inside a scripted collaborative-learning session changes group chat dynamics | (c) groups often adopt AI content (incl. verbatim copying); no significant gain in group-answer quality - **Caveats:** 3-page short paper; Wizard-of-Oz (human-relayed ChatGPT output); unequal cell sizes (e.g. 57 vs 16 group submissions); participant n not reported; student sample - **Screening check:** meets criteria (education exception) — genAI takes part in a structured group session (Pyramid collaboration script) and group-level outcomes (group submissions, group chat) are measured; method and sample (submissions) reported; English. #### R90 — GenAI and team communication in secondary education - **Citation:** Hu, W., & Chan, C. K. Y. (2026). How Generative AI Affects Team Communication in Secondary Education: A Qualitative Investigation. Journal of Computer Assisted Learning, 42(4). https://doi.org/10.1002/jcal.70270 - **Type:** journal — peer-reviewed: yes - **Full text read:** abstract only (Wiley PDF blocked by Cloudflare; article number/pages not verified) - **Method:** mixed-methods/qualitative: written reflections from all participants, semi-structured interviews; thematic analysis (RQ1) and activity-theory analysis (RQ2) (abstract) - **n:** 24 secondary students (Grades 9-12) from two Hong Kong schools; 10 interviewed (abstract) - **Setting:** student teams using GenAI tools during group tasks in secondary school; specific tools not verified - **Findings:** 1. GenAI improved perceived communication quality by fostering deeper understanding, generating new ideas and resolving disagreements (abstract). 2. GenAI "significantly" reduced communication frequency among team members and could introduce delays in team exchanges (abstract; "significantly" is the authors' wording in a qualitative study). 3. Six factors shaped GenAI use in team communication: perceived value of discussions, task interdependence, team familiarity, attitudes towards AI, guidelines for AI use, role clarity (abstract). - **Relevance:** (a) — | (b) — | (c) GenAI may enrich idea generation and disagreement resolution in teams while reducing member-to-member communication - **Caveats:** abstract only; self-reported student perceptions; education setting — AI is used by teams during group tasks rather than facilitating them. - **Screening check:** borderline education case — included because GenAI takes part in structured group work and a group-level process (team communication) is examined; if the review applies the education rule strictly (AI must facilitate/participate in the session), this could be EX-EDU. #### R560 — Generative AI in children's creative collaboration - **Citation:** Yoo, D., Newman, M., Pitt, C. R., Vo, K. H., Zhang, M., Kim, M., et al. (2026). Generative AI in Children's Creative Collaboration: Impact, Perception, and Design Guidelines. Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26), ACM, pp. 1-23. https://doi.org/10.1145/3772318.3790907 - **Type:** conference — peer-reviewed: yes - **Full text read:** abstract only (ACM DL blocked by Cloudflare; no preprint found) - **Method:** five co-design sessions with children using GenAI tools; focus group interviews with parents; qualitative analysis (abstract) - **n:** 28 children aged 5-11 across 5 co-design sessions; 17 parents (abstract) - **Setting:** children collaborating creatively with diverse GenAI tools (text, image, video, voice); specific tools not verified - **Findings:** 1. GenAI fostered positive social dynamics by enabling "Human vs. AI" teaming and co-creation with shared ownership (abstract). 2. GenAI disrupted collaboration when roles between children were unclear, when the AI ignored group dialogue and when the AI dominated children's agency (abstract). 3. Children and parents envisioned a socially attuned AI in an "older sibling" role that scaffolds while allowing playful disagreement; parents raised overreliance concerns (abstract). - **Relevance:** (a) — | (b) conditions under which GenAI helps or disrupts a group creative session (role clarity, attending to group dialogue) | (c) GenAI can support or dominate group co-creation - **Caveats:** abstract only; children; co-design (participatory research) setting; no quantitative outcomes reported in abstract; group sizes per session not verified. - **Screening check:** meets criteria — empirical qualitative study, groups of children collaborating with GenAI in structured co-design sessions (not classroom), method and n stated, English. #### C243 — Scoping review: generative AI and student group work - **Citation:** Wei, Y., & Perkins, M. (2026). Generative AI and student collaboration: a scoping review of group work processes, outcomes, and risks. International Journal for Educational Integrity, 22(1), Article 8. https://doi.org/10.1007/s40979-026-00217-x - **Type:** review (scoping review, PRISMA-ScR) — peer-reviewed: yes - **Full text read:** partly. The Springer HTML was read through WebFetch as a model summary, not verbatim, and page numbers were unavailable because the PDF download was blocked. The details below come from the abstract and that summary. - **Method:** Scoping review following PRISMA-ScR. Scopus, Web of Science and Google Scholar were searched for English-language publications from January 2023 to March 2025, and included studies were analysed with reflexive thematic analysis. - **n:** 18 included studies (higher education, with one exception involving professionals in ideation workshops, per the WebFetch summary, not verified). - **Setting:** Higher-education student group work with ChatGPT and other LLM tools. - **Findings:** 1. Reported benefits: group knowledge development, idea generation, reflective thinking support, communication efficiency, task coordination and timely feedback (abstract). 2. Reported risks: reduced peer-to-peer interaction and critical engagement when tools are over-relied on, and privacy, transparency and bias concerns (abstract). Outcomes depend on task design, integration into the group workflow and AI literacy (abstract). 3. Eight themes, including "GenAI mediates communication" and "GenAI transforms team interaction". Some studies found participants felt "less collaborative" and that efficiency gains came with reduced communication demand (WebFetch summary, not verbatim; not verified). - **Relevance:** (a) — | (b) AI in coordinating and mediating student group work (secondary evidence) | (c) idea-generation benefits versus reduced peer interaction and critical engagement - **Caveats:** Secondary review, not primary data. Student context. The full-text details are unverified (model summary). Useful for backward citation search. - **Screening check:** FAILS: secondary scoping review, not an empirical primary study. Use it as a source for snowballing rather than as an included study. #### C332 — Group conversational agents: systematic review - **Citation:** Yeo, S., Zhang, T., Bateman, S., Hsieh, G., Kim, Y.-H., Perrault, S. T., Li, J., & Tang, A. (2026). Group Conversational Agents: A Review of Designs that Support and Shape Group Interaction. Proceedings of the 2026 ACM Designing Interactive Systems Conference (DIS '26), pp. 2765–2779. https://doi.org/10.1145/3800645.3812934 - **Type:** conference (systematic review) — peer-reviewed: yes - **Full text read:** yes (author copy, https://younghokim.net/files/papers/yeo-group-ca-dis26.pdf; page numbers below are PDF pages of that copy, which still carries a template header) - **Method:** Systematic, design-oriented review following an adapted PRISMA framework; primary search in ACM Digital Library and IEEE Xplore for English peer-reviewed studies 2020–2025, plus backward/forward snowballing via Google Scholar (which also surfaced pre-2020 studies); inclusion required natural-language agent, ≥2 humans interacting with it, group-level intervention, group awareness, synchronous interaction, empirical grounding; descriptive coding (challenge, agent role, construction as rule-based / Wizard-of-Oz / LLM-based, modality, study method, group size, relationship) by two lead authors (§3, pp. 3–5; Table 1, p. 6). - **n:** 53 peer-reviewed studies (publication years in Table 2 range 2000–2025) (§3.1.2, p. 5; Table 2, p. 7). Screening counts per PRISMA stage are in Figure 1 but not verified from the extracted text. - **Setting:** Conversational agents (chatbots, robots, avatars; text or voice) in multi-party human groups: meetings, deliberations, group chats, learning groups, teams. - **Findings:** 1. The corpus is organised around four group challenges that agents target: orientation (engagement/participation), conflict (mediating disagreement), alignment (structuring communication, coordination, shared understanding) and execution (ongoing task support) (§1, p. 2). 2. Agents are framed in three recurring roles (facilitator/mediator, peer/team member, background assistant); facilitator/mediator interventions are mostly delivered publicly to the group, positioning the agent as "a neutral party with some authority" (§5, pp. 11–12; Table 4, p. 13). 3. GCAs are "predominantly designed as short-term interventions targeting specific group challenges within bounded interactional contexts", mostly evaluated in lab studies; structural tensions include visibility vs. discretion, proactivity vs. group autonomy, and agent authority vs. group ownership (§5, p. 11; §6, pp. 14–15; abstract). - **Relevance:** (a) facilitator use of AI: — | (b) AI planning/running/supporting group sessions: maps design space and evidence base of agents acting as facilitators/mediators/moderators in group sessions | (c) AI and group ideation/decisions: identifies conflict/alignment interventions (e.g., LLM devil's advocate) but reports no pooled outcome data - **Caveats:** Descriptive, not evaluative synthesis (no effect sizes); corpus mixes rule-based, Wizard-of-Oz and LLM-based agents back to 2000 — the number of LLM/generative-AI studies could not be verified from the garbled Table 2; overlaps with primary studies already in this review (e.g., Chiang et al. LLM devil's advocate; Kim et al. moderator chatbot). - **Screening check:** Meets criteria as a systematic review (method and n = 53 reported, English, group-level conversational agents including LLM-based ones); borderline on the generative-AI criterion because much of the corpus is pre-LLM. ## Topic-eligible but not assessable These records look eligible on their title and abstract, but the method, the sample size, the group format or the use of generative AI could not be verified, because the full text was paywalled or blocked and the abstract does not say. They are not counted as included and should not be quoted beyond their abstracts. With institutional access, several would likely move into the included set. | Key | Title | Year | DOI/URL | Why not assessable | |---|---|---|---|---| | OA256 | Generative AI as a Socio-Technical Facilitator: A Case Study of AI-Supported Group Formation and Co-Creation | 2026 | https://doi.org/10.1109/mipro70003.2026.11591640 | no abstract or full text retrievable | | OA275 | LLM-based Automated Facilitator for Building Effective Consensus on Mission and Vision Definition | 2024 | https://doi.org/10.1109/ica63002.2024.00034 | closed access; abstract reports no sample size | | OA293 | Development of a Generative AI Facilitator Supporting Creative Problem-Solving and Thinking Flow Analysis: Focused on Informal Learning Contexts | 2025 | https://doi.org/10.36358/jce.2025.25.2.1 | closed access; paper language (possibly Korean) and design not verifiable | | OA364 | Towards Group Decision Support with LLM-based Meeting Analysis | 2025 | https://doi.org/10.1145/3708319.3733646 | abstract only; no sample size reported | | OA392 | Leveraging Generative AI to Address Unproductive Silences in Virtual Meetings | 2025 | https://doi.org/10.2139/ssrn.5702194 | no abstract or full text accessible | | OA412 | Assessing LLMs for Prioritization in Meeting-Based Group Recommendations | 2025 | https://doi.org/10.1109/ictai66417.2025.00050 | abstract only; no sample size | | OA451 | Cueing the Crowd: LLM-Driven Conversational Cues Across Different Meeting Modalities Increase Topical Diversity of Generated Ideas | 2025 | https://doi.org/10.1145/3715928.3737486 | no abstract or full text accessible | | OA509 | Applying generative AI to explore new dimensions of reflection and interaction in agile design workshops | 2025 | https://doi.org/10.1504/ijasm.2025.144170 | abstract only; participant count not reported | | OA519 | LLM-Based Structuring of Oral Discussion in Workshop to Support Collaboration Among Local Government and Simulated Citizens | 2024 | https://doi.org/10.1007/978-3-031-67998-8_1 | abstract only; no sample size | | OA523 | Just Talk, and Sticky Notes will be Created: Towards Collaborative Dialogue in Face-to-Face Workshops the Experiment with a Generative AI-Based | 2025 | https://doi.org/10.1145/3706599.3720231 | abstract only; no sample size | | OA764 | Advancing Knowledge Together: Integrating Large Language Model-based Conversational AI in Small Group Collaborative Learning | 2024 | https://doi.org/10.1145/3613905.3650868 | abstract only; no sample size | | OA767 | Enhancing student engagement in online collaborative writing through a generative AI-based conversational agent | 2024 | https://doi.org/10.1016/j.iheduc.2024.100979 | no verbatim abstract or full text; n only from search snippets | | OA925 | Design Futures with GAI: Exploring the Potential of Generative AI Tools in Collaborative Speculation | 2023 | https://doi.org/10.1007/978-3-031-48057-7_10 | abstract only; no sample size, group format unclear | | OA931 | Stimulating Brainstorming Activities with Generative AI in Higher Education | 2025 | https://doi.org/10.1109/educon62633.2025.11016340 | full text embargoed; n not verifiable | | OA939 | How the integration of generative AI in design thinking-based collaborative learning impacts student creativity | 2026 | https://doi.org/10.1080/14606925.2026.2652387 | abstract only; sample size not stated | | OA1346 | Serendipity Wall: A Discussion Support System Using Real-time Speech Recognition and Large Language Model | 2024 | https://doi.org/10.1145/3652920.3652931 | abstract only; no sample size | | OA1365 | Examining the Effect of ChatGPT on Small Group Ideation Discussions | 2024 | https://doi.org/10.1007/978-3-031-59373-4_12 | abstract only; no sample size | | OA1468 | Designing AI for team ideation: How content strategy and participation style affect creative performance through team emergent states | 2025 | https://doi.org/10.1016/j.ergon.2025.103824 | no abstract or full text accessible | | OA1567 | The Role of Generative Artificial Intelligence in a Brainstorming Session: A Feasibility Study | 2026 | https://doi.org/10.1007/978-3-032-26235-6_10 | abstract only; n not stated; possible overlap with Gerardini et al. 2025 | | OA2312 | The Effects of an AI Participant on Online Group Deliberation: Evidence from a Controlled Study with Afghan Youth | 2026 | https://doi.org/10.1007/978-981-95-9667-6_5 | abstract only; no sample size | | OA2423 | Can AI support inquiry, co-creation, and integration functions in Citizens’ Assemblies? | 2026 | https://doi.org/10.1177/02633957261466242 | abstract only; participant numbers not reported | | S2-520 | Image Generative AI to Design Public Spaces: a Reflection of How AI Could Improve Co-Design of Public Parks | 2024 | https://doi.org/10.1145/3656588 | Full text inaccessible (ACM Cloudflare block) and abstract gives no sample size; case study/reflection using Dream Studio as design probe on interview data | | S2-527 | Enabling Partial Participation in Remote Meetings | 2026 | https://doi.org/10.1145/3772318.3790744 | abstract only; whether live group meetings were involved not verifiable | | S2-547 | Projecting Residents’ Intentions: Participatory Interior Renovation Design through AI Image-Generation Workshops | 2026 | https://doi.org/10.1051/shsconf/202623302004 | Full text not retrievable (bot protection) and abstract gives no sample size; AI image-generation participatory workshops with residents | | C61 | Co-Ideation Across Dimensions: The Effects of Mediation Mode on Creative Collaboration and Ai Co-Ideation | 2025 | https://doi.org/10.2139/ssrn.5381161 | no abstract retrievable | | C62 | HarmonizeAI: Enhancing Group Dynamics Through AI-Assisted Collaborative Music Generation | 2025 | https://doi.org/10.1109/aixheart65685.2025.00011 | sample size not verifiable | | C173 | Externalizing Intent: Preserving Semantic Alignment in Graphic Recording Generation from Dialogue | 2026 | https://doi.org/10.1145/3772363.3798935 | no abstract retrievable | | C342 | Collaborative Decision-Making in Human-AI Versus Human-Human Teams: The Influence of Information Inquiry and Team Performance Expectations on Team Pro | 2026 | https://doi.org/10.1177/00187208261477411 | abstract only; unclear whether human-AI teams had >=2 humans | | C371 | TeamWise: Exploring Virtually Embodied AI Facilitation for Video-Based Team Onboarding | 2026 | https://doi.org/10.1145/3772363.3798791 | sample size not verifiable | | C404 | From Interactive Idea Development to Solution Planning: An AI-assisted Discussion Facilitation Timeline for Collaborative Planning | 2024 | https://doi.org/10.1145/3657054.3659120 | unclear whether the chatbot is generative | | C548 | Generative AI as a Mediator in Creator Collaboration: Challenges, Design Opportunities, and Concerns | 2026 | https://doi.org/10.1145/3816940 | abstract only; unclear whether sessions were joint | | C633 | Neutral Voice in Echo Chambers: Effects of AI Participation on Group Polarization and Social Perception | 2026 | https://doi.org/10.1145/3772363.3798386 | abstract only; total sample size not stated | | C639 | Institutionalising AI into Democratic Deliberations: Tensions in Practice | 2026 | https://doi.org/10.1007/978-3-032-36676-4_10 | no abstract retrievable | | C734 | MedAIting with the Mediators: Exploring Aligned Agent Interventions to Resolve Challenges in Mediator Workflows | 2026 | https://doi.org/10.1145/3805689.3806726 | no abstract; only web summary available | | C764 | When design workshops meet chatbots: Meaningful participation at scale? | 2024 | https://doi.org/10.1177/14780771241253440 | abstract only; no sample size | | C783 | Exploring collaborative decision-making: A quasi-experimental study of human and Generative AI interaction | 2024 | https://doi.org/10.1016/j.techsoc.2024.102662 | full text blocked; sample size not stated in abstract | | C822 | Ideenfindung und Konzeptgestaltung in der Produktentwicklung mittels generativer künstlicher Intelligenz | 2026 | https://doi.org/10.1007/978-3-658-49371-4_13 | German chapter; no abstract retrievable | | R4 | Dittos: Personalized, Embodied Agents That Participate in Meetings When You Are Unavailable | 2024 | https://doi.org/10.1145/3687033 | Meetings with embodied agent representing absent member (WoZ, 10 groups n=39; focus groups n=24); abstract does not state generative AI and full text blocked (ACM Cloudflare) - generative AI involvement not verifiable | | R6 | Relational AI: Facilitating Intergroup Cooperation with Socially Aware Conversational Support | 2025 | https://doi.org/10.1145/3706598.3713757 | No abstract in OpenAlex/Crossref; ACM full text blocked; no preprint found | | R10 | Investigating the Impact of Generative Artificial Intelligence on Brainstorming: A Preliminary Study | 2023 | https://doi.org/10.1109/icce-taiwan58799.2023.10226617 | Gen-AI brainstorming experiment; abstract gives no sample size and does not clarify group vs individual co-ideation; full text (IEEE) not accessible | | R40 | AI-Facilitated Collective Judgements | 2025 | https://doi.org/10.2139/ssrn.5167340 | No abstract available; SSRN full text blocked; study type unknown | | R82 | Augmenting Online Teamwork Dynamics: Investigating the Effects of a Generative AI Chatbot on Perceived Assistance and the Assessment of Team Members | 2024 | https://doi.org/10.1080/10447318.2024.2400406 | Generative AI chatbot in online student team chat (3 conditions); abstract gives no sample size and full text (Taylor & Francis) not accessible | | R83 | Chatbots in Collaborative Settings and their Impact on Virtual Teamwork | 2025 | https://doi.org/10.1145/3710945 | Chatbot assistant in 4-person virtual teams (72 participants, 18 teams); abstract does not state the chatbot is generative/LLM and ACM full text blocked - generative AI involvement not verifiable | | R142 | When AI Joins the Negotiation Table: Evaluating AI as a Moderator | 2025 | https://doi.org/10.1145/3719160.3736630 | No abstract available; ACM full text blocked; no preprint found | | R179 | Quieter Learning, Safer Teams: Human Role Plasticity in Hybrid Human-AI Teams | 2025 | https://doi.org/10.2139/ssrn.5754222 | AI teammate experiment (134 teams) and field survey (307); abstract does not state generative AI and SSRN full text blocked - generative AI involvement not verifiable | | R206 | Robots in the Room: A Qualitative Exploration of Democratic Agency and AI-Augmented Deliberations | 2026 | https://doi.org/10.1145/3776734.3794554 | Humanoid robot in six small-group deliberations (4 humans + robot each); abstract does not state whether the robot used generative AI/LLM and full text not accessible - generative AI involvement not verifiable | | R209 | Artificial Influence: How AI Role and Gender Reshape Team Dynamics | 2026 | https://doi.org/10.2139/ssrn.6469423 | AI advisor/assistant in 3-person VR investment teams; abstract does not state generative AI and SSRN full text blocked - generative AI involvement not verifiable; abstract gives no n | | R218 | Enhancing Consensus-Building Feedback Through Psycholinguistic and Epistemic Augmentations With Large Language Models | 2026 | https://doi.org/10.1109/tsmc.2026.3693564 | LLM-generated persuasive feedback integrated in a fuzzy consensus model for group decision-making; abstract gives no sample size and does not state whether human decision-makers took part; IEEE full text not accessible | | R239 | Can AI Representatives Be Legitimate? An Exploratory Study in Citizen Participation | 2026 | https://doi.org/10.1007/978-3-032-36676-4_11 | No abstract available; ZORA/Springer full text blocked (bot check) | | R290 | Social perception in Human-AI teams: Warmth and competence predict receptivity to AI teammates | 2023 | https://doi.org/10.1016/j.chb.2023.107765 | No abstract available; Elsevier full text not accessible | | R304 | When and What to Speak: An Autonomous Embodied Chat Agent in Symmetrical Reality | 2026 | https://doi.org/10.1109/vrw70859.2026.00241 | No abstract available; IEEE 2-page poster not accessible | | R693 | Human‐ AI collaboration: Designing artificial agents to facilitate socially shared regulation among learners | 2024 | https://doi.org/10.1111/bjet.13534 | Wizard-of-Oz prototype metacognitive agent for collaborating learners; abstract gives no sample size and full text (Wiley) blocked; WoZ design also makes generative AI involvement doubtful | | R726 | Evaluating an AI Documentation Assistant for Anesthesiology Teams | 2025 | https://doi.org/10.1145/3706599.3706658 | No abstract available; ACM full text blocked | | R850 | How to Use Generative AI to Support Collective Problem-Solving | 2025 | https://doi.org/10.2139/ssrn.5335953 | No abstract in OpenAlex/Crossref; SSRN full text blocked (403); design and n unknown | | R942 | Students’ agency in GenAI-mediated group assessment: an ecological-emergent perspective | 2026 | https://doi.org/10.1080/02602938.2026.2730644 | No abstract available; T&F full text blocked; design/n and AI role in group sessions unknown | | R955 | Artificial Intelligence Augmenting Human Teams. A Systematic Literature Review on the Opportunities and Concerns | 2022 | https://doi.org/10.1007/978-3-031-05643-7_4 | Systematic review 2022; no abstract available and Springer full text blocked; number of studies and genAI coverage unknown | | R967 | Arguing with AI: LLMs as a Debate Partner in the Classroom | 2025 | https://doi.org/10.1109/tale66047.2025.11346727 | No abstract available; IEEE full text blocked; whether debates are group sessions and n unknown | | R979 | Effects of chatbot-assisted in-class debates on students’ argumentation skills and task motivation | 2023 | https://doi.org/10.1016/j.compedu.2023.104862 | No abstract available; ScienceDirect blocked; whether the chatbot is generative/LLM-based and n unknown | | R1056 | Mapping the Horizon: A Scoping Review of Generative AI in Collaborative Learning | 2025 | https://doi.org/10.1109/tale66047.2025.11346681 | Scoping review of GenAI in collaborative learning; no abstract, IEEE blocked; method/number of studies unknown | | R1060 | Designing an AI Speculative Facilitator for Healthcare Futures: From Prototype to Participatory Deployment | 2026 | https://doi.org/10.1007/978-3-032-30080-5_1 | AI speculative facilitator deployed participatorily (title); no abstract, Springer blocked; n unknown | | R1080 | Exploring the Supportive Role of Artificial Intelligence in Participatory Design: A Systematic Review | 2024 | https://doi.org/10.1145/3661455.3669868 | review read from abstract only; generative-AI scope not confirmable | | R1114 | Assisting Group Discussions Using Desktop Robot Haru | 2024 | https://doi.org/10.1109/icra57147.2024.10611098 | Robot Haru assists group discussions (dialogue, balance, gaze modules); abstract does not state whether dialogue is LLM/generative; IEEE full text blocked, genAI criterion unverifiable | | R1224 | Impact of AI-agent-supported collaborative learning on the learning outcomes of University programming courses | 2025 | https://doi.org/10.1007/s10639-025-13487-8 | No abstract available; Springer blocked; design/n and AI role unknown | | R1507 | How generative AI supports novice project managers in creative reasoning: evidence from a pilot experiment in risk management | 2026 | https://doi.org/10.1108/ijmpb-10-2025-0451 | AI-assisted vs non-AI teams compiling a risk register; abstract gives no sample size and Emerald full text inaccessible | | R1625 | Artificial Intelligence Tools Enhance Design Thinking in Multidisciplinary Design and Education – A Systematic Review | 2026 | https://doi.org/10.1007/978-3-032-29178-3_14 | No abstract available; Springer blocked; number of studies and genAI coverage unknown | | R1629 | Human–GenAI Interaction in Collaborative Learning Environments: A Scaffolding Perspective in Higher Education | 2026 | https://doi.org/10.1109/itet70052.2026.11650872 | No abstract available; IEEE blocked; design/n unknown | | GS3 | Simpson, J., Patil, G., Stening, H., Bin Kamruddin, A., Somerville, D., Seage, S., Nalepka, P., Dras, M., Hosking, S. G., Kallen, R. W., Richardson, M | | https://doi.org/10.1016/j.chbah.2026.100278 | full text blocked; no abstract with sample size | | GS5 | Wang, Y., Chen, H., Wei, X., Chang, C., & Zuo, X. (2025). Trusting the machine: Exploring participant perceptions of AI-driven summaries in virtual fo | | https://doi.org/10.1016/j.chbah.2025.100198 | abstract only; no sample size; group format unclear | | GS6 | Chen, H., Wang, Y., & Chang, C. (2025). Enhancing qualitative inquiry: AI-assisted focus group data collection. Qualitative Research Journal, advance | | https://doi.org/10.1108/QRJ-04-2025-0145 | abstract only; no sample size | ## Adjacent evidence excluded under the protocol The protocol excludes studies of individual AI use that say nothing about groups, and keeps only Doshi & Hauser (2024) as the reference case for group-level effects. The following studies were excluded on that rule, but they test the same mechanism (AI-assisted individuals producing less diverse ideas as a collective, or an AI replacing a human partner). They are listed here so a protocol amendment could add them if needed: - OA1418: Human diversity fuels collective creativity that large language models cannot simulate or sustain (2026). https://doi.org/10.48550/arxiv.2607.26899 — individual writing; collective-diversity measure only (adjacent to Doshi & Hauser) - OA1627: Diverse AI personas can mitigate the homogenization effect in human-AI collaborative ideation (2026). https://doi.org/10.1016/j.chbah.2026.100289 — individual writing; collective-diversity measure (adjacent) - OA1633: How AI Ideas Affect the Creativity, Diversity, and Evolution of Human Ideas: Evidence From a Large, Dynamic Experiment (2024). https://doi.org/10.48550/arxiv.2401.13481 — individual ideation; collective-diversity measure (adjacent) - C100: Homogenization Effects of Large Language Models on Human Creative Ideation (2024). https://doi.org/10.1145/3635636.3656204 — individual ideation; collective-diversity measure (adjacent) - OA1649: Creative or uncreative partner: Comparing humans and AI in collaborative creative tasks (2026). https://doi.org/10.1016/j.chb.2026.109172 — human-human vs human-AI pairs (adjacent) - C468: “Who” Is the Best Creative Thinking Partner? An Experimental Investigation of Human–Human, Human–Internet, and Human– AI Co‐Creation (2024). https://doi.org/10.1002/jocb.1519 — human-human vs human-AI dyads (adjacent) - C543: Two-Player Alternate Uses Test: A Controlled Testbed for Interactive Human-AI and Human-Human Co-Creation (2026). https://doi.org/10.1145/3803784.3816851 — two-player human-AI vs human-human (adjacent) - C817: Inspiration booster or creative fixation? The dual mechanisms of LLMs in shaping individual creativity in tasks of different complexity (2025). https://doi.org/10.1057/s41599-025-05867-9 — human vs LLM partner (adjacent) - OA1326: Brainstorming with a Generative Language Model: Effect of Exposure to AI Ideas on Brainstorming Performance and Cognitive Load (2025). https://doi.org/10.1007/s12599-025-00974-y — human-AI dyads vs solitary humans; no multi-human group - OA2047: Wisdom of the Crowd, Without the Crowd: A Socratic LLM for Asynchronous Deliberation on Perspectivist Data (2025). https://doi.org/10.1145/3757707 — LLM replaces group deliberation partner (adjacent) - GS12: Ju, H., & Aral, S. (2025). Collaborating with AI Agents: A Field Experiment on Teamwork, Productivity, and Performance. arXiv:2503.18238. (). https://arxiv.org/abs/2503.18238 — Compares human-human pairs vs human-AI pairs; AI condition is a human-AI dyad, no multi-human group with AI - OA1314: Homogenizing Effect of a Large Language Model (LLM) on Creative Diversity: An Empirical Comparison of Human and ChatGPT Writing (2024). https://doi.org/10.31234/osf.io/8p9wu — human vs GPT-4 essays (AI-only output) - C22: Towards Collaborative Brain-storming among Humans and AI Agents: An Implementation of the IBIS-based Brainstorming Support System with Multiple AI Agents (2024). https://doi.org/10.1145/3643562.3672609 — In AI condition single individuals brainstorm with GPT agents; no multi-human group with AI ## PRISMA-style count | Source | Records retrieved | Duplicates across queries of the same source | Duplicates of records from an earlier source or other versions of one work | Screened (title/abstract) | Excluded at title | Excluded after abstract/full text | Not assessable | Included | |---|---|---|---|---|---|---|---|---| | OpenAlex | 3,033 | 91 | 516 | 2,426 | 2,170 | 182 | 21 | 53 | | Semantic Scholar | 1,584 | 59 | 893 | 632 | 596 | 15 | 3 | 18 | | OpenAlex citation check (round 1) | 826 | 0 | 10 | 816 | 613 | 146 | 13 | 44 | | OpenAlex citation check (round 2) | 1,727 | 0 | 26 | 1,701 | 1,596 | 57 | 29 | 19 | | Google Scholar cross-check (web search) | 15 | 0 | 0 | 15 | 0 | 7 | 3 | 5 | | **Total** | **7,185** | **150** | **1,445** | **5,590** | **4,975** | **407** | **69** | **139** | Citation-check counts are records not already retrieved from the databases (published 2022 or later). Rows in the Semantic Scholar source that were already retrieved from OpenAlex are counted as duplicates. Excluded after abstract or full-text screening, by reason: | Code | Reason | Records | |---|---|---| | EX-NOTEMP | not empirical (opinion, concept, framework or prototype without evaluation, narrative review, protocol, podcast) | 106 | | EX-INDIV | individual human–AI use, no group of two or more humans (includes human–AI dyads) | 87 | | EX-NOGENAI | AI involved is not generative (rule-based agents, classifiers, IBIS/D-Agree agents, robots without an LLM, Wizard-of-Oz without a generative system) | 66 | | EX-TOPIC | off topic, or AI used only as a research analysis tool | 61 | | EX-SIM | LLM agents or simulated humans only | 26 | | EX-EDU | classroom use of AI without a structured AI-supported group session | 24 | | EX-NLP | offline NLP benchmark on transcripts not about facilitation | 15 | | EX-NOTSTUDY | not a study (proceedings, dataset, supplementary material, replication package, workshop or panel description) | 11 | | EX-LANG | not in English or German (Korean, Japanese, Spanish) | 6 | | EX-NOSAMPLE | no sample size or method reported in the accessible text | 5 | | | **Total** | **407** | Title-screen exclusions (EX-TITLE): 4,975. Duplicates and other versions removed (EX-DUP plus cross-query duplicates): 1,595. Not assessable: 69. **Included: 139** (four of them are reviews: OA1894, C243, C332, R1658). What this search covers: the listed databases, queries and citation rounds on 2026-09-26. It does not cover every eligible study that exists. Known gaps: eleven Semantic Scholar relevance queries could not run because of rate limits (partly offset by the Boolean bulk queries); the Google Scholar step was a web-search cross-check rather than a systematic 50-results-per-query screen; the broad OpenAlex queries were narrowed to title-level or relevance-ranked subsets; and 69 records could not be assessed without subscription access.