Anthropic has built an open source system designed to verify which artificial intelligence model generated a particular output, according to a blog post published Tuesday. The system, named FlashProve, offers what researchers describe as "provable inference" — a cryptographic proof that functions as a mathematical guarantee confirming a specific output originated from a particular model. It's intended to allow users to confirm the answer they got actually came from the model they believe they're using.
The system operates by requiring AI labs to release a "fingerprint" for each model they launch, which can subsequently be matched to particular outputs and confirmed cryptographically. Importantly, these fingerprints don't expose a model's weights — the internal parameters that determine how it functions and which closed-model developers like Anthropic and OpenAI keep secret. That allows labs to share the fingerprints without surrendering their most valuable intellectual property, making the system compatible with commercial models as well. The system is also engineered to be lightweight, so confirmation could be done on a personal device. As part of the reveal, researchers released a comprehensive technical specification and functional code for different components of the system, which is publicly accessible under an open source license on GitHub.
Much like cryptographic email signatures that allow recipients to verify a message genuinely originated from its stated sender, FlashProve would be resistant to imposters trying to pass off one model as another, even when the imposter is the laboratory that created and operates the model. The researchers note that the system could be deployed to stop an AI company from quietly substituting a less expensive model to reduce expenses, or a pricier model to exaggerate its performance metrics. FlashProve could also be employed to authenticate claims by outside parties that a particular task was completed by an advanced model. "A developer might claim that their software library was scanned for vulnerabilities by a frontier cybersecurity model, or a company might claim that its business plan was given the go-ahead by an appropriate model," the researchers write. "How can a user verify these claims?"
Work on the system continues, and no plans exist to implement it at Anthropic or any other laboratory. "This research post is meant to share intermediate progress, and not a completed research agenda," researchers explain in the announcement post. "We are sharing this explicitly with the goal that others look at what we have done, take it, and improve it." The release reflects growing concerns about transparency and trust as AI systems become more widely deployed across sensitive applications, from cybersecurity scanning to business planning. By enabling verification without compromising proprietary model architectures, the approach attempts to balance commercial interests with the need for accountability in an industry where users often can't independently confirm what system they're actually interacting with. Labs adopting such verification could face pressure from competitors who prefer opacity, while enterprises evaluating AI tools may soon demand cryptographic proof as a standard procurement requirement.