New Proofs Enhance AI Safety Without Needing Debate.

As AI models become increasingly powerful, ensuring their outputs align with human intentions is paramount for safety. A common approach, known as "debate," involves two competing AI models arguing their case to convince a weaker human verifier. However, this model assumes equal AI capabilities and the truthfulness of one AI, which may not always be realistic in practice. This new work explores an alternative: "single-prover" interactive proofs for AI safety, aiming to bypass the need for a debate. Traditional single-prover proofs often fall short when computations involve external information sources, such as human input or web databases. This research addresses this gap by developing "doubly-efficient" single-prover interactive proofs and arguments specifically for oracle-aided computations. The proposed method is effective in scenarios where the computation is robust (meaning minor oracle errors don't change the outcome) or when the oracle behaves like a low-degree polynomial. These findings suggest that reliable interactive verification of AI outputs is achievable even without the complexities and assumptions of a debate framework, especially under specific conditions of oracle access.