opML (Optimistic Machine Learning) is a method for running machine-learning inference for blockchain applications in which results are accepted as correct by default and can be overturned only if someone proves, during a set challenge period, that they are wrong. It lets a smart contract use the output of an AI model without having to run the model on-chain or generate an expensive cryptographic proof for every result.
The approach borrows the fraud-proof design of an optimistic rollup and applies it to AI. It was set out in a January 2024 research paper by a team at Hyper Oracle, the project now known as ORA, and is usually discussed as the cheaper counterpart of zkML. Like zkML, opML is an early-stage technique with a small number of live deployments, and its security depends on at least one honest party watching for errors.
opML in plain terms
Picture a community noticeboard where anyone can pin up a result, such as "the model says this photo shows a cat". Nobody checks it straight away; it simply stays up for a week. During that week, anyone who thinks it is wrong can raise a dispute. If they are right, the person who posted the false result loses the deposit they put down, and the challenger is rewarded. If nobody objects, the result becomes final.
This works because cheating is costly and only one honest watcher is needed to catch it. The trade-off is patience: you have to wait out the objection period before treating the answer as settled, and everything on the noticeboard is public.
How opML works
The opML paper describes a system with submitters, who run models and post results, and verifiers or challengers, who re-run the same model to check them. The main steps are:
- A requester, often a contract, asks for an inference, such as a text or image generated by a specified model.
- A submitter runs the model off-chain on ordinary hardware and posts the result with a bond, a deposit similar to staking.
- During a challenge window, any verifier can re-run the computation. Because results must be reproducible, the design uses fixed-point arithmetic and software floating-point libraries so that every honest machine gets the same answer.
- If a verifier disagrees, the two sides play a bisection game, repeatedly splitting the computation in half until they isolate a single disputed step.
- That one step is executed on-chain by a fraud-proof virtual machine, which settles who was right. The loser's bond is slashed.
To avoid compiling an entire large model into the dispute machine, the paper proposes a multi-phase protocol: disputes are first narrowed at the level of the model's computation graph, and only the final small piece is checked at instruction level. The authors describe the security as an "any-trust" guarantee: a single honest verifier is enough to force a correct outcome.
opML compared with zkML
| Feature | opML | zkML |
|---|---|---|
| Verification method | Fraud proofs after a challenge | Validity proofs for every result |
| Cost per inference | Low; ordinary hardware | High proving cost |
| Model size | Can handle large models, including language models | Practical mainly for small and medium models as of 2026 |
| Finality | Delayed until the challenge window closes | Immediate once the proof is checked |
| Privacy | None; inputs and model must be public to be re-run | Can hide inputs or model weights |
The same research group later proposed "opp/ai", a hybrid that proves privacy-sensitive parts of a model with zero-knowledge techniques and handles the rest optimistically. A related idea for data feeds is the zkOracle, which uses zero-knowledge proofs instead of fraud proofs.
Uses
opML is aimed at cases where AI output must be usable on-chain but the model is too large to prove directly:
- On-chain AI requests. ORA has offered an on-chain AI oracle that lets contracts on Ethereum and other networks request outputs from open-source models, with results secured optimistically.
- AI-driven applications. Games, NFT projects using non-fungible tokens and prediction tools can call models without trusting a single server.
- Agents and automation. AI agents that trigger transactions can have their model calls made disputable, adding a check on automated behaviour.
Limitations
opML inherits the weaknesses of optimistic systems. If no honest verifier is online during the challenge window, a wrong result can become final, so the scheme depends on watchers having incentives to keep checking. The waiting period makes it unsuitable for decisions that must settle instantly, such as liquidations in DeFi. Because verifiers must re-run the model, inputs and weights cannot stay private, which rules out many uses involving personal or proprietary data. Deterministic execution of AI models across different hardware is also hard to guarantee in practice. Like any blockchain oracle, an opML service can only vouch that a computation was done correctly, not that the model itself gives good answers, and AI outputs carry their own bias and error risks, especially where large sums or leveraged positions depend on them.
The approach is part of a wider effort to make Layer 2 style verification work for general computing, and it uses the same economic logic of bonds and penalties for misbehaving validators that secure proof-of-stake networks.