Anthropic develops FlashProve to verify AI model identity

Anthropic is developing an open-source system called FlashProve that could verify which AI model generated a particular response. The project uses what its researchers call “provable inference”: a cryptographic proof intended to provide a mathematical guarantee that an output came from a specified model.

The system is designed to resist attempts to present one model as another, including by the company operating the model. It could help detect the undisclosed use of a cheaper model to reduce costs or a more expensive one to improve reported performance. It could also let users check third-party claims that an advanced model performed a task, such as reviewing software for vulnerabilities or assessing a business plan.

For FlashProve to work, AI developers would publish a fingerprint for each model. The fingerprint could be matched with outputs without exposing the model’s weights, allowing commercial providers to participate while keeping those internal parameters private. Verification is intended to run on a local device. Anthropic researchers released a technical specification and working code on GitHub under an open-source license. The research remains in progress, and Anthropic has no plans to deploy the system at present.