Anthropic releases FlashProve research for AI output verification

Anthropic publishes an open design for verifying model outputs

Anthropic has published FlashProve, an open-source research system intended to prove cryptographically which AI model generated a particular output. The project addresses what its researchers call “provable inference”: a mathematical guarantee that an answer originated from a specified model rather than an impersonator or an undisclosed substitute.

The company released a detailed technical specification and working code for parts of the system under an open-source licence on GitHub. Anthropic stresses that the work is still in progress and that it has no plans to deploy FlashProve at Anthropic or at another AI laboratory.

The proposed mechanism resembles a cryptographic email signature. A recipient can verify that a message came from its claimed sender; FlashProve is designed to let a user verify that an AI response came from the model named by a provider. The check is intended to remain reliable even if the impersonator is the laboratory that built and operates the model.

Model fingerprints without releasing weights

For FlashProve to operate, AI labs would publish a fingerprint for every released model. That fingerprint could later be matched to an output and checked cryptographically. Anthropic says the fingerprint would not disclose the model weights, the internal values that shape model behaviour and remain private for closed-model providers.

This separation is central to the proposal. Labs could make model identity testable while retaining their weights, allowing the approach to be considered for commercial as well as openly released models. Anthropic also designed verification to be lightweight enough to run on a local device.

The initiative arrives as the market continues to divide between open and proprietary approaches, with the coexistence of open and proprietary AI models illustrating how those two paths can coexist while model provenance becomes more relevant. FlashProve does not establish which model is best for a task; it is intended to establish whether the stated model was actually used.

Why verifiable inference could matter

Anthropic identifies several situations in which a model provider’s assertion may need independent confirmation. An AI company could quietly replace an advertised model with a cheaper one to reduce costs, or use a more expensive model to make its performance look stronger than the deployed service. In either case, a conventional user may have little way to distinguish the claimed model from the one behind the output.

The same issue extends to third-party claims. A developer might say a frontier cybersecurity model scanned a software library for vulnerabilities, while a company might say an appropriate model reviewed its business plan. FlashProve is intended to give users a way to test the identity claim attached to such work.

For businesses, the practical implication is that AI model selection and governance may increasingly require evidence of the model used for consequential outputs, particularly where security checks, approvals or reported results depend on a named system.