Anthropic is developing FlashProve, an open-source system designed to let users cryptographically verify which AI model produced a particular output. The approach, known as provable inference, is intended to provide mathematical evidence connecting a response to a specific model.
Like a digital signature, the system could help confirm that a model's identity matches the one claimed. It is designed to work even when the service provider operates the model, and could support independent checks of claims that a particular AI system was used for a task.
Model fingerprints without revealing weights
FlashProve relies on AI developers publishing a fingerprint for each model. That fingerprint can later be checked against outputs, but does not expose the model's weights--the internal parameters that shape its behavior. Anthropic says this could allow the method to work with commercial models while keeping their underlying technology private.
The system is intended to be lightweight, with verification possible on a local device. Anthropic has shared a technical specification and working code under an open-source license, inviting others to assess and improve the approach. The company says the research is ongoing and FlashProve is not currently planned for deployment.
If developed further, verifiable model identities could make AI services easier to audit and help users assess whether a system's capabilities match the claims made about it.
