A research team argues that making AI models argue with each other is the wrong way to keep AI honest.
The paper, posted to arXiv, proposes "disagreement resolution" as a replacement for the standard debate framework used in scalable oversight. In the debate setup, AI agents argue opposing positions to help a judge reach the correct answer — but the authors identify a core flaw: models are rewarded for being persuasive, not for being right. Their alternative borrows from human mediation practice, directing models to collaboratively pinpoint where they disagree, examine the evidence together, and either reach consensus or isolate the precise sticking point. The result is a measurable improvement: non-expert models hit 62.1% judging accuracy under disagreement resolution versus 49.2% under standard debate.
The gap matters because scalable oversight is one of the few credible frameworks researchers have for supervising AI systems that may eventually surpass human judgment in specialized domains. If the dominant method systematically rewards rhetoric over truth, that is a structural problem — not a tuning problem. The mediation framing is a genuine reorientation, not an incremental patch.
Debate as an oversight mechanism has attracted serious attention from labs including Anthropic and OpenAI; this result does not retire the approach, but it does give the field a sharper benchmark to beat.