Anthropic develops cryptographic proof for AI model identity

The mechanism functions by requiring labs to publish a unique model fingerprint, allowing users to verify outputs without exposing the model's proprietary internal weights. By applying principles similar to digital email signatures, the system prevents companies from surreptitiously switching between models to manipulate performance metrics or reduce operational costs. It also enables third parties to authenticate claims regarding which specific AI handled a given task, such as a security audit or high-stakes business analysis.

Designed for efficiency, the verification process can run on local hardware, ensuring accessibility for end-users. While the researchers have released a technical specification and open-source code on GitHub, the project remains in the experimental phase. Anthropic currently has no immediate plans to deploy the technology, explicitly positioning the release as a collaborative invitation for the broader research community to refine and expand upon the framework.