A new AI debate protocol makes honesty the dominant strategy
This protocol gives human overseers stronger guarantees when two AIs debate problems too complex for direct review: honest, correct answers are a dominant strategy for both debaters. It applies to problems with sufficiently stable decompositions, where a hard question can be broken into simpler claims, and improves on prior work by proving worst-case correctness instead of average-case correctness. The authors also prove no black-box protocol can do better on any individual instance.