Anthropic is developing FlashProve, a system designed to verify whether a response was generated by the AI model a company says it uses. It relies on cryptography to produce a mathematical proof linking an output to a specific model.
The proposal could help users check that they are interacting with the advertised AI rather than a different or less capable system. Anthropic says the approach could also make it harder for providers to secretly switch to cheaper models or claim performance results based on a model they did not use.
FlashProve uses what the researchers call “provable inference.” Model developers would publish a digital fingerprint for each system, which could then be checked against its outputs using cryptographic methods. The fingerprints would not disclose a model’s weights, the internal values that shape its behavior, and verification is designed to run on a local device.
The system could be relevant beyond direct AI services. For example, a developer might want to verify a claim that an advanced cybersecurity model was used to inspect a software library for vulnerabilities. A company could also seek to confirm that an appropriate AI model reviewed its business plan.
Anthropic presented the project in a company publication on Tuesday, November 3, and made a detailed technical specification and parts of the code available under an open-source license. The company has not announced plans to implement FlashProve in its own services or on other AI platforms. Its researchers described the release as an intermediate step in the research, inviting other teams to examine and improve the proposal.
Sources: Olhar Digital.