New Anthropic Prototype Will Help Verify Authenticity of AI Responses

Anthropic has released FlashProve, a system for verifiable AI inference. It is designed to confirm that a response was generated by the stated model rather than another one covertly substituted to reduce costs or alter test results.

For the system to function, developers must publish a cryptographic 'fingerprint' of the model. This does not expose the model's weights, so the approach can be applied to commercial closed models as well. Anthropic has also made the technical documentation and source code available under the Apache 2.0 license. The research is ongoing, and there are currently no plans to implement FlashProve at Anthropic or other labs.