AI Devil's Advocate
AI Devil's Advocate is a decision check in which a team asks an AI chat assistant to play a friendly dissenting teammate and question a decision the team has made or is about to make. The assistant asks one question at a time about alternative views, drawbacks, evidence and assumptions, and the team then decides what, if anything, to change. The technique comes from one of the 'AI as teammate' prompts in Ethan and Lilach Mollick's 2023 paper 'Assigning AI: Seven Approaches for Students, with Prompts', written for student teams and adapted here for workshop use.
Facilitation script
- 1
Explain the purpose: the team will let an AI assistant argue against its decision to find weak spots, and will judge every objection itself. Have the team write the decision and its main reasons in a few sentences.
5 min - 2
Enter the role prompt and the decision. Agree who types and remind the group that answers are given by the team, out loud, one question at a time.
3 min - 3
Work through the assistant's questions on alternative viewpoints and drawbacks. Mark any question the team cannot answer.
8 min - 4
Continue with the questions on evidence and assumptions. If the chat turns repetitive, ask the assistant for the strongest objection it has not raised yet, then stop.
7 min - 5
Review the exchange away from the screen. Sort the challenges into valid and new, already handled, and wrong or generic.
5 min - 6
Decide what changes in the decision and who follows up on open questions. Note in one or two sentences how the challenge affected the outcome.
5 min
Tips
Keep the typing and the thinking separate: the person at the keyboard should record the team's answer, not their own.
Treat the assistant's objections as prompts for the team's judgement, not as findings, because it can be confidently wrong or simply generic.
The most useful moment is usually a question the team cannot answer, so mark those and follow them up with real data.
Close the laptop for the final decision step so the humans own the conclusion.
Common pitfalls
Deferring to the assistant because its objections sound fluent. The team changes a sound decision on the strength of a generic argument
Accepting invented facts. The assistant may cite figures or cases that do not exist, and unchecked they end up in the decision record
Letting one person type and answer alone, which turns a team check into a private chat and leaves the rest of the group uncommitted to the result
Describing the decision in one vague line, so the assistant can only return textbook objections that apply to any decision
Ending when the chat ends, without a human review step. The session then produces a transcript and no change to the decision
Variations
With larger groups, split into teams of three to six, give every team the same decision and the same role prompt, and compare which objections each chat produced. For a remote session, one person shares their screen while the others answer by voice. A stricter variant has each person write their own doubts silently before the chat starts, so the team can see which concerns the assistant raised that nobody in the room had. The original paper presents the prompt for classroom teams and gives no timing or group size; the figures here are a workshop adaptation.
Where it fits
When to use it
A team reached agreement quickly and nobody argued the other side
A decision is drafted but not yet announced, so there is still room to amend it
The team is homogeneous or senior voices dominate, and dissent from a colleague would carry a social cost
You want a low-stakes rehearsal of the objections stakeholders are likely to raise
A small team has no outsider available to review its reasoning
When not to use it
The decision involves confidential, personal or client data that may not be entered into the AI tool your organisation allows. Run a human devil's advocate round or Six Thinking Hats instead
The team is in open conflict about the decision. An AI challenger adds noise to a disagreement that needs a facilitated conversation between the people involved
You need failure scenarios for a concrete plan. A Premortem produces more specific and better grounded risks
The decision turns on specialist facts the assistant cannot know, such as internal figures or local regulation. Ask a subject expert
The decision is final and cannot be changed, in which case challenging it only produces frustration
Related methods
Frequently asked questions
What is AI Devil's Advocate?▾
AI Devil's Advocate is a short decision check in which a team asks an AI chat assistant to act as a dissenting teammate. The assistant questions a decision one point at a time, asking about alternatives, drawbacks, evidence and assumptions. The team answers, reviews the challenges critically and decides whether the decision needs to change.
Where does the AI Devil's Advocate technique come from?▾
It is based on a prompt in the 'AI as teammate' section of 'Assigning AI: Seven Approaches for Students, with Prompts' by Ethan Mollick and Lilach Mollick (2023). The paper was written for student teams and does not specify timing or group size. The workshop format described here is an adaptation of that prompt for facilitated team sessions.
How is this different from a premortem?▾
A premortem asks the team to imagine that a plan has failed and to explain why, which produces failure scenarios from the team's own knowledge. AI Devil's Advocate brings in an outside questioner that challenges the reasoning behind a decision. The two combine well: use the devil's advocate on the choice itself and the premortem on the plan that follows from it.
Can the AI be trusted to find the real weaknesses?▾
No, and the method does not depend on that. The assistant knows only what the team tells it, can state wrong facts with confidence and often raises generic objections. Its value is in asking questions that no team member wants to ask a colleague, so every challenge has to be judged by the people who know the context.
How many people can take part?▾
Three to six people per chat works well, because everyone can still contribute to the answers. Larger groups should split into parallel teams, each with its own device, and compare results afterwards. One shared screen with twenty people watching turns the exercise into a demonstration.
Plan your next workshop with AI
Workshop Weaver helps you combine methods like AI Devil's Advocate into a complete, timed agenda in minutes.
Try it freeMethod descriptions on Workshop Weaver are original content written by our team, based on established facilitation practices. This method was inspired by work from Ethan Mollick and Lilach Mollick, 'Assigning AI: Seven Approaches for Students, with Prompts' (2023).