FlashProve: Anthropic explores proof of AI model identity

Anthropic is working on FlashProve, a system intended to establish that a response really comes from a particular AI model. According to TechCrunch, which reports on a publication dated November 3, the project relies on cryptographic proof of the model’s execution. One of its aims is to allow users to verify the identity of the system they believe they are interacting with.

The mechanism requires laboratories to publish a fingerprint of their models that can then be associated with responses and verified cryptographically. According to the reported presentation, this fingerprint does not reveal the model’s weights. The system is also designed to allow verification on a local device. These remain objectives and characteristics described by the researchers, with no independent demonstration presented in the article.

One proposed use would be to detect the undisclosed substitution of one model for another: a cheaper model to reduce expenses, or a more expensive one to inflate reported performance results. FlashProve could also be used to check a provider’s claim that a task was performed by a specific model.

The aim would then be to make part of the relationship between provider and customer verifiable. Proof of identity alone would not, however, be enough to establish the correctness or safety of a response: those are separate questions from provenance.

The project’s maturity remains limited. The researchers have published a technical specification and code for various components under an open source license. They describe this work as an intermediate step. No deployment at Anthropic or another laboratory is planned, according to TechCrunch.