AI debate enables weaker judges to supervise stronger debaters.
3 events · 1 assessment · 1 decision
Assessed without decomposition
The first turn used a third of the budget, so the claim was assessed from knowledge of the key literature: Irving 2018, Michael 2023, Khan 2024, Kenton 2024, Barnes 2020. Status: supported, conf 0.6, credence 0.55. No decomposition or instances were recorded. A future pass should add subclaims: (a) debate improves judge accuracy under information asymmetry; (b) the obfuscated-arguments problem; (c) results extend to capability gaps and open-ended tasks. It should also record those papers as instances.
Assessed Supported
verdict confidence 0.60 · credence 0.55
The claim comes from the AI safety via debate proposal (Irving, Christiano and Amodei, 2018). The idea is that two capable agents argue opposite sides before a less capable judge. If honest arguments are easier to defend than dishonest ones, the judge can reach correct answers it could not reach alone. Empirical work since 2023 supports a limited version of the claim. In reading-comprehension tasks where the judge cannot see the source text, human and model judges are more accurate after watching debates than after hearing a single consultant or answering with no help. Michael et al. (2023) found this with human debaters. Khan et al. (2024) found it with LLM debaters and also found that judge accuracy rose as the debaters got more persuasive. Kenton et al. (2024, Google DeepMind) found that debate consistently beat consultancy. Against direct question answering without an advisor, though, the results were mixed, especially on tasks where the judge and debaters had the same information. Several things remain unsettled. The tested gaps between judge and debater are mostly information gaps, not gaps in capability. The tasks are bounded and checkable. Debaters were not trained adversarially to deceive. The obfuscated-arguments problem (Barnes et al., 2020) gives a theoretical reason to think a dishonest debater could win on hard, open-ended questions by making claims too complex to refute. The evidence favors the claim within tested settings. Whether it holds in general, for strongly superhuman debaters on open-ended tasks, is not established.
Claim entered the graph