Anthropic Develops Open-Source FlashProve to Verify AI Model Identity

Anthropic has released FlashProve, an open-source system designed to address the challenge of users being unable to confirm which specific model generated an AI response. In this system, model providers first publish an encrypted "fingerprint" corresponding to their model weights, then generate proofs for specific input-output pairs. Individuals holding the fingerprint, input, and response can verify the proofs without accessing the model weights or re-running the model. Researchers believe this "provable inference" can help users identify if service providers have covertly switched models and can be used to verify third-party claims, such as statements that code has been checked using a specific model. Anthropic has published a technical report, protocol documentation, and open-source code to allow external review and improvement. However, proof generation is not cost-free. In tests conducted using a large open-source model, each proof required approximately 19 MB of data, and generating proofs consumed GPU time equivalent to 70 to 95 times that of standard inference; verification, however, can be performed on devices without GPUs. Anthropic emphasized that this is an interim research result and has not announced plans to deploy FlashProve within its Claude services. The system verifies whether a specified model produced a corresponding output, but does not guarantee the correctness of the answer itself.