Key answer
Statements by several experts written by one AI are not independent pieces of evidence. They can provide different approaches, but “three experts agree” is not equivalent to support from three independent sources.
Full explanation
This does not make multi-role discussion useless. NAACL 2024's Solo Performance Prompting had one model generate personas, discuss and revise solutions across three task types; GPT-4's logic-grid score rose from 57.7% to 68.3%. The full method also included two demonstrations and iteration, so the improvement cannot all be attributed to a greater number of experts. Methods and Table 2
Du et al.'s multiagent debate instead had multiple model instances answer first, then exchange opinions, improving some reasoning and factual tasks. This differs from staging a roundtable within one output. Agreement itself still does not establish truth. Study
The reasons are concrete: later “experts” in one generation have already seen earlier statements. Even separate calls to the same model can share knowledge gaps. More perspectives and more independent evidence are different things.
One useful request is:
Review this proposal separately for implementation cost, user experience and failure risks. For each perspective, give the strongest objections, evidence and items needing validation. Do not require consensus.
This helps find blind spots. Verify disputes through the cited material or actual test results rather than counting votes.