Anthropic has released technical specifications and code for FlashProve, a system designed to let users verify which AI model generated a specific output. According to Anthropic’s announcement, FlashProve works by using cryptographic proofs, similar to email signatures, to guarantee that an output comes from a particular model. This could prevent AI providers from secretly swapping models to cut costs or inflate performance claims.
The system requires AI labs to publish a unique fingerprint for each model, which can then be matched to outputs without exposing proprietary model weights. This approach means commercial labs can participate without risking their intellectual property. Verification is designed to be lightweight, allowing checks to be performed locally.
For regulated sectors, the ability to confirm a model’s identity supports audit trails and compliance with governance requirements. However, Anthropic says FlashProve is still in the research stage and there are no current plans to deploy it in production.
