Digital Watermarks: Anthropic Builds Open Source System to Verify AI Model Identities

Anthropic announced work on a new open-source system designed to verify which exact artificial intelligence model generated a specific text output. The system, named FlashProve, provides cryptographic proof that confirms the origin of generated content. Known in computer science research as provable inference, the process gives users a mathematical guarantee that the answers they receive come directly from the specific model they intended to use.

FlashProve functions like digital signatures on emails, letting recipients verify that a message originated from the claimed sender. The verification protocol stops labs from disguising one model as another. Researchers noted that FlashProve prevents developers from swapping in cheaper, low-power models to cut compute costs or routing queries to high-end systems temporarily to inflate benchmark results. By anchoring outputs to verifiable cryptographic proofs, users can double-check model claims independently.

Beyond verifying model identity, FlashProve helps companies validate third-party work claims. Software teams can prove that an advanced cybersecurity model audited a codebase for vulnerabilities, or business founders can confirm that a specific frontier model vetted a financial plan. Cryptographic verification gives organizations clear audit trails when relying on automated software output for critical operational choices.

To run the verification process, artificial intelligence developers publish a digital fingerprint for every released model. The verification system matches specific text outputs against the published fingerprint using cryptographic algorithms. Digital fingerprints protect proprietary intellectual property because they do not reveal model weights, parameters, or internal training data. As a result, commercial developers like Anthropic and OpenAI can share model fingerprints publicly without exposing trade secrets. Additionally, the lightweight system allows users to verify model proofs locally on personal computers.

Anthropic published a technical specification document alongside working open-source code on GitHub, making the framework available for public review and testing. The company explained that research remains ongoing across the industry, with no immediate plans to deploy FlashProve commercially across Anthropic platforms or external services.

Anthropic researchers emphasized that the published code represents early progress rather than a finished product. By releasing the project open-source, the lab encourages external developers, security researchers, and competing platforms to build upon the codebase. Sharing open protocols helps establish transparent standards for model verification as automated systems handle higher volumes of sensitive digital work.

Developing reliable verification standards becomes necessary as synthetic text fills public communication channels. Without clear verification tools, bad actors can spoof model names, misrepresent software capabilities, or fake automated safety audits. Building open-source cryptographic frameworks helps secure digital workflows, ensuring that organizations can trust the identity and integrity of their automated tools.