Anthropic, the company behind the Claude AI models, has published early research on FlashProve, an open-source system designed to prove which AI model produced a given answer. It works like the digital signatures that let an email recipient confirm a message really came from its sender. The aim is to let anyone check that a response came from the model it claims to come from.

It’s research for now: Anthropic says it has no plans to deploy it, at Anthropic or at any other lab.

The idea is called provable inference, a mathematical proof that a specific model produced a specific output. For it to work, an AI company would publish a "fingerprint" for each model it releases, and any answer could later be checked against it. Crucially, the fingerprint does not reveal the model's weights, the internal numbers that make it work and that companies like Anthropic and OpenAI keep secret.

That means closed commercial models could take part without giving away their most valuable asset. The system is also light enough to run the check on an ordinary device.

Anthropic's researchers say it is built to hold up even against the company running the model. That covers a lab quietly swapping in a cheaper model to save money, or a more expensive one to make its benchmark results look better. It could also be used to check claims made by others.

"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?" Anthropic has released a technical specification and working code on GitHub under an open-source licence. "We are sharing this explicitly with the goal that others look at what we have done, take it, and improve it," the researchers say.

TechCrunch's Russell Brandom reported the research and noted that it is shared as intermediate progress rather than a finished system. Anthropic's research post sets out how the system works, and the code is on GitHub for anyone to inspect and build on.

Most people using AI have no way to check which model actually answered them. A business paying for a premium model through an app or reseller simply has to trust that it is getting that model and not a cheaper one. That trust has been tested before, as when Meta was caught presenting a specially tuned version of Llama 4 on a public leaderboard, which we noted when covering Meta's Muse Wonder.

FlashProve answers a different question from watermarks. A watermark, like the one OpenAI now adds to ChatGPT text in the EU, shows that AI was involved. This would show which AI. As more vendors sell products built on AI and make claims about the models behind them, a way to check those claims would be useful to any business buying AI services.

Whether it happens depends on rival labs agreeing to publish fingerprints, so treat this as a promising idea rather than something you can use yet.

You might also like our piece on AI agent swarms in Minecraft, Anthropic's IPO filing, and AI-built software clones.

Get more SA tech and business news and subscribe to The Open Letter.