A debate-based oversight system reaches 95% agreement with human expert panels
Researchers deployed a scalable-oversight system in which two models argue opposing sides of a safety-critical decision while a smaller judge model evaluates the debate — reaching 95% agreement with human expert panels. It is a route to overseeing systems too complex or fast for humans to check directly.
Debate is an answer to the scalability problem in oversight. As models tackle decisions too complex or numerous for humans to review each one, you need a way to surface the reasoning for judgment — and having two models argue opposing sides forces the considerations into the open, where a lighter-weight judge can adjudicate. It makes a hard decision legible enough to check.
The 95% agreement with expert panels is the number that makes it credible. Matching human experts nearly all the time suggests the debate structure is extracting the right considerations, not just producing confident noise — the difference between a plausible mechanism and one you might actually trust to oversee decisions at a scale humans cannot.
The honest limit is the remaining 5% and the judge's own reliability. A system that agrees with experts 95% of the time disagrees 5% of the time, and on safety-critical decisions the tail is where the danger lives. Debate oversight is a genuine step toward scalable checking, but it moves trust onto the judge model and the debate structure, which themselves need scrutiny.
Claude 5 Hub — AI safety 2026: alignment research breakthroughs → · Medium — ICLR 2026 oral papers in AI safety: a 35-paper deep dive →