On this page
What blockchains still can't do with AI
The most-hyped quadrant — AI as the rules of a protocol, or a chain that “runs AI” — is the one to be most skeptical of, for a structural reason: blockchains execute deterministic computation that every node repeats, and running a neural network that way is absurdly expensive. A model inference that costs milliseconds and fractions of a cent on a GPU would burn enormous for every verifier. So the honest architecture questions are all about trusting what a model produced off-chain: it is the again, one level up. Who attests that “the model really said that” — the operator, a committee, a TEE, a cryptographic proof?
Why the model runs off-chain
The research answer is : generate a that a specific model, on specific inputs, produced a specific output — so a contract can verify the AI's claim without re-running the model. It builds on the same machinery as ZK rollups (prove a computation instead of repeating it), and early demos exist for small models. The open problems are size (proofs for large models remain expensive), model secrecy (a model you prove against may have to be revealed), and adversarial inputs (a proof says the model ran honestly — not that the input wasn't crafted to make it misbehave). Until those shrink, any “on-chain AI” product is doing one of two things: calling an off-chain model and trusting the caller, or putting an oracle in the middle — both of which are fine, as long as the marketing admits it.