Anthropic is developing an open-source project called FlashProve, designed to provide provable inference for artificial intelligence models. The system uses cryptographic proofs to verify the authenticity of an AI model's output, ensuring that users are interacting with the specific model they expect.
The technology acts as a digital signature for AI responses, preventing companies from secretly substituting models to reduce operational costs or manipulate performance benchmarks. It remains effective even if the AI developer attempts to conceal the substitution.
FlashProve allows third parties to verify claims about AI-driven tasks, such as security audits or business planning. Developers publish a digital fingerprint of each model, which is cryptographically matched to specific outputs without exposing proprietary model weights.
The system is designed to be lightweight, enabling verification directly on a user's local device. Currently, the project is in the research phase with no immediate plans for commercial deployment by Anthropic or other laboratories.
Anthropic has released the technical specifications and source code for the system components under an open license on GitHub. The company encourages the research community to review, adopt, and improve upon the current progress.
