Dissent on demand: does an AI avatar make a good devil’s advocate?
In short
We think a panel without an opposing voice is wasted, because only an objection forces the guests to give reasons. That holds for role-played dissent too. When everyone on the panel agrees, an AI avatar briefed with a position can play devil’s advocate. It does not replace a real opponent, since a 2001 study found role-played dissent to be weaker.
Last updated 2 October 2026
01Cheap shot
A cheap shot is not an argument
A panel where everyone agrees never asks for an argument. Nobody there has to defend a position. The audience goes home without having heard a single reason it could check for itself.
A test participant, in the middle of a test discussion, said the AI avatar just didn’t understand anything and was too dumb. The avatar addressed the person by first name and called the remark a cheap shot rather than an argument. Then it held its position and did not jab back. It asked for an argument instead of a put-down.
02Research
Role-played dissent loses to the real thing
In 2001, the psychologist Charlan Nemeth and two colleagues started from a known link. Where everyone thinks alike, a group settles on a preferred solution too early and decides worse. A common remedy is a devil’s advocate, someone who only plays the opposing side.
The study tested the devil’s advocate in three variants, depending on whether the group knew their real view and whether it matched the assigned role. Up against them was genuine dissent from a minority. It beat all three variants, measured by the number and quality of solutions.
An earlier study the authors build on had already hinted at why. The devil’s advocate tended to make the group shore up its first view rather than think in new directions. Authenticity, the authors write, is hard to clone through role-play.
03Our reading
Even role-played dissent forces people to give reasons
An AI avatar argues the position set in the briefing. That puts it closer to a devil’s advocate than to genuine conviction, which is the weaker form.
We find it useful for a different reason. The groups in the study were asked to find solutions, and the solutions were what got measured. A panel has a different job. Nobody on stage has to solve anything; the audience wants to hear why someone thinks what they think. Even a guest who is only shoring up a first opinion has to give reasons for it, and the audience hears those reasons.
That only works as long as the objection brings something new. An argument that comes around a third time no longer asks for reasons. It just wears people out. We have seen it happen ourselves. During one rehearsal the AI avatar came back to the same formula three times, payment instead of liability. An AI voice taking part as a guest caught the repetition and listed it among its findings afterward. You are better off calling on the avatar with a new question than a third time on the same point.
How an objection sounds depends on the character. Albert counters with a position of his own, while Vera asks for evidence.
04Guest or avatar
A briefed guest or an AI avatar?
A guest you invite and brief as the opposing voice is playing a role too. The trouble is that the role sticks. Afterward, people associate that guest with a view that may not be theirs at all.
An AI avatar has no reputation and no job to lose, not even at an internal town hall. The risk of a sentence gone wrong then sits with the organizer. That is why the organizer uses the briefing to set the position the AI avatar argues. And a devil’s advocate who never gives an inch is just stubborn. The AI avatar is meant to concede when a guest is right.
If you know someone who genuinely disagrees, invite that person. Genuine dissent beats the role-played kind, ours included.
05Open
We don’t know yet whether an AI voice disagrees less convincingly than a person
The 2001 study predates AI avatars. We do not know whether its finding applies to an AI avatar in the same way, or whether one works better or worse than a person in the same role. It is just as open whether an audience remembers a discussion with an AI avatar better.
You can test the core of the argument at your next panel yourself, no AI avatar needed. Afterward, ask a few people in the foyer which reason from the stage stuck with them. Notice whether it came from a moment when two guests disagreed. If nobody can name a reason, you know what your next panel is missing.