AI Safety via Debate

Debate trains agents to argue competing answers before a human judge. The proposal aims to make difficult reasoning easier to check by focusing disagreement on smaller, assessable claims.^[1]

It is a form of Scalable Oversight, not a guarantee that the most persuasive answer is true. The 2020 research writeup investigates protocols using human experts and non-expert judges.^[2]

The 2020 work reports unreliable early judging and ambiguity failures; cross-examination was a proposed response, with unresolved training concerns.^[2]

See the original proposal and the human protocol experiments.

Evidence and error localization

The original image experiment used machine search and a fixed classifier; its results do not measure human judging of unrestricted language.^[1] The later human writeup asks whether debaters can actually locate an error small enough to expose. A capable judge still needs usable evidence from the debate.^[2]