Having two systems argue opposite sides of a question in front of a human judge, on the theory that it is easier to spot the flaw in an argument than to work out the answer alone. It is one of the more concrete proposals for supervising systems that know more than the supervisor.
Technical view
A zero-sum game where models argue for competing answers and a judge, human or model, decides, with the hope that truthful positions are easier to defend. Theoretical work relates the setting to complexity classes and to obfuscated arguments, where a dishonest debater constructs a case too complex to refute within the depth available. Empirical results are mixed and sensitive to judge capability, debate protocol and topic.
Related
Part of AI Terminology Index, section: Alignment and Safety Concepts.