Verifiable AI inference

When a smart contract, a user or another agent relies on an AI answer, it usually has no way to know which model produced it or whether the operator tampered with the result. Verifiable inference adds evidence. The three main approaches are zero-knowledge proofs of the computation, known as zkML; trusted execution environments, where secure hardware signs an attestation of the code that ran; and optimistic or consensus methods, where several parties run the model and disputes are checked by re-execution.

The goal matters most where AI outputs move money or decide outcomes: oracles that resolve prediction markets, agents trading on behalf of users, credit scoring in lending protocols or content moderation in decentralized apps. Several networks and rollups now advertise verifiable inference as a core feature.

Each method trades off differently. Zero-knowledge proofs give strong guarantees but are slow and costly for large models; secure enclaves are fast but depend on trusting the chip maker and have suffered side-channel attacks; re-execution needs deterministic results, which large models do not always produce on different hardware. Proving which model ran also says nothing about whether its answer is correct.