Checking an AI answer is never free, it just charges you in a different currency each way.
a) Heavy cryptographic proof (zkML) charges compute.
b) Trusted-hardware approaches (TEEs) charge a dependency on the chip maker.
c) Consensus-scoring charges accuracy, because it cannot actually prove the model ran.
Ours charges a small overhead per response, a company claim we have not had independently benchmarked, running on testnet today. Every one of these is a real cost, and the only wrong move is pretending your option is free.