Paying for an advanced AI model requires an uncomfortable amount of faith. You send a prompt, an answer appears, and you simply trust that the vendor did not quietly swap in a cheaper model to cut costs. On Tuesday, Anthropic published a blog post detailing FlashProve, an open-source system built to replace that blind trust with a mathematical guarantee.
The Old Bait and Switch
Researchers call the concept "provable inference." It works like a cryptographic email signature that confirms a message really came from the claimed sender, letting users verify that an answer actually came from the model they think they are using.
What makes FlashProve interesting is who it treats as the suspect. The system guards against impostors disguising one model as another, even when the impostor is the very AI lab that built and runs the model.
Incentives to cheat run in both directions. An AI company might secretly swap in a cheaper model to cut costs, or slip in a more expensive model to inflate its performance results.
Third parties make unverified boasts, too. "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?"
Fingerprints Without the Weights
Verifying a commercial model is tricky when the company refuses to show you what is inside it. FlashProve works around that secrecy through a few practical design choices:
- Public model fingerprints: AI labs publish a cryptographic fingerprint for each model they release, which can later be matched to specific outputs and verified cryptographically.
- Protected internal weights: The fingerprints do not reveal a model's weights, the internal numbers that define how it behaves. Closed-model makers like Anthropic and OpenAI keep those numbers private, so labs can publish fingerprints for commercial models without giving up their most valuable asset.
- Local device checks: Researchers designed the system to stay lightweight so verification can run on a local device.
- Open source code: Alongside the announcement, the team released a detailed technical specification and working code for various elements of the system under an open-source license on GitHub.
A Lie Detector Waiting for Volunteers
Writing the math is the easy part. Getting commercial AI labs to willingly lock themselves into cryptographic honesty is another matter entirely.
Research into FlashProve is ongoing, and neither Anthropic nor any other lab has plans to deploy it right now. "This research post is meant to share intermediate progress, and not a completed research agenda," the 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."
Anthropic has handed the industry the working blueprints for a polygraph test. Now everyone gets to watch who volunteers to strap in first.