AI debate systematically favors the debater arguing for the true answer over the debater arguing for a false one.

The claim is the premise on which AI safety via debate rests, often put as "it is harder to lie than to refute a lie": when two AI systems argue opposite answers before a judge, the side arguing for the true answer should win more often. In the settings tested so far it holds. Human-judged debates on reading-comprehension questions (Michael et al., 2023), LLM debates judged by weaker models or humans without access to the text (Khan et al., 2024), and a broad study of weak LLM judges over strong debaters (Kenton et al., 2024) all found the correct side winning clearly more often than chance, and debate beating single-advocate consultancy. Khan et al. also found that making debaters more persuasive through inference-time optimization raised judge accuracy, and later work training debaters with reinforcement learning on the same kind of task reported the same direction, which partly answers the worry that only prompted, honesty-tuned debaters were tested.

The evidence is narrow. Most of it comes from tasks with an information asymmetry and a checkable quotation from a hidden text, where the true side can point to verifiable evidence; on tasks without that asymmetry Kenton et al. found smaller or mixed advantages. The obfuscated-arguments problem identified by Barnes and Christiano (2020) shows a structural route by which a dishonest debater could win on hard problems: by making a long argument whose flaw neither the honest debater nor the judge can locate efficiently. Later protocol work (such as prover-estimator debate) aims to blunt this, but under assumptions not yet tested at scale. Judges, human and model, are also known to respond to confidence and rhetoric unrelated to correctness.

The claim is therefore supported as a description of current experimental results and open as a general property of debate between much stronger, fully trained debaters on open-ended questions. What would settle it is evidence from debaters trained adversarially to win on tasks without verifiable quotations, at capability levels well beyond the judge.