Anthropic is developing an open system called FlashProve designed to verify exactly which artificial intelligence model generated a specific result. The project implements the concept of so-called “provable inference”: the user receives cryptographic proof and a mathematical guarantee of the data source authenticity, confirming that they are indeed working with the claimed model rather than a substitute. The initiative has been published in a research format, and the specification and working code of the system elements are hosted under an open license on GitHub.
What is “Provable Inference”
Much like cryptographic email signatures, FlashProve provides protection against attempts to pass off one model as another. A key property of the system is that it remains effective even if the laboratory that developed and deployed the AI attempts to conceal the substitution. Researchers note that the technology can be used to prevent situations where an AI company secretly deploys a cheaper model to cut costs or, conversely, an expensive one to artificially inflate benchmark scores.
Verifying Third-Party Claims
FlashProve can also be used to verify third-party claims that a particular task was solved using advanced neural networks. “A developer might claim their software library was audited for vulnerabilities by a state-of-the-art cybersecurity model, or a company might claim their business plan was approved by a corresponding model. [Yet] how can a user verify these claims?”, the project authors write. Thus, the system transforms marketing statements into a verifiable cryptographic fact.
“Digital Fingerprint” Without Revealing Weights
For verification to work, AI laboratories must publish a “digital fingerprint” of each released model, which is subsequently cryptographically matched against specific responses. Importantly, these fingerprints do not reveal model weights—the internal numerical parameters determining AI behavior that closed commercial system developers like Anthropic and OpenAI keep secret. Consequently, laboratories can publish fingerprints without risking their intellectual property. The system itself is designed to be lightweight, enabling verification directly on the user's local device.
Project Status and Open License
Currently, FlashProve is in the research stage: there are no immediate plans for deployment by Anthropic or other laboratories. The detailed specification and working code of the system elements are published under an open license on GitHub. “This research publication is intended to share interim progress rather than a finished research agenda. We are explicitly sharing this so that others can study what we have done, build upon it, and improve it,” the project authors explained. The initiative could become a vital trust tool in the generative AI ecosystem, where the authenticity of the output source is becoming an increasingly significant factor for businesses and regulators.
