Anthropic Builds FlashProve to Verify Which AI Model Answered

What happened

Anthropic is building an open-source system that can prove which AI model produced a given answer. The company described the system, called FlashProve, in a blog post published Tuesday. It is designed to give users what researchers call "provable inference": a cryptographic proof, essentially a mathematical guarantee, that an output came from a specific model.

The idea is simple to state and hard to fake. A user who thinks they are talking to one model should be able to check that claim, rather than take a company's word for it. FlashProve works in a way similar to cryptographic email signatures, which let a recipient confirm that a message really came from its claimed sender. Anthropic says the system would hold up even against an impostor that is the lab that built and runs the model.

That matters because model swaps are otherwise invisible to users. Researchers say a company could secretly swap in a cheaper model to cut costs, or a pricier one to inflate its performance results, and the user would have no way to tell. FlashProve is meant to close that gap. It could also verify claims by third parties that a specific task was done by an advanced model. The researchers give two examples: a developer claiming a software library was scanned for vulnerabilities by a frontier cybersecurity model, and a company claiming its business plan was given the go-ahead by an appropriate model.

Why it matters

Trust is becoming the weak point of AI, not capability. Users now rely on models for code review, research, medical-adjacent questions, and business decisions, yet they cannot see which system actually did the work. A proof that ties an answer to a named model would turn a marketing claim into something checkable. That is a shift from trusting a brand to verifying a result, and it fits a wider push for labels and accountability around AI-generated content. Pennsylvania, for example, has moved to [require labels on AI-generated ads](/article/pennsylvania-require-labels-ai-generated-ads), a different tool aimed at the same problem: people should know what they are looking at.

The design tries to solve the obvious objection. Labs would publish a "fingerprint" of each model they release, which can later be matched to specific outputs and verified cryptographically. Those fingerprints do not reveal the model's weights, the internal numbers that define how it behaves and that closed-model makers like Anthropic and OpenAI keep private. Labs could therefore publish fingerprints without giving up their most valuable asset, so the system could work with commercial models too. Anthropic also says the system is lightweight, so verification could run on a local device rather than requiring a trusted server.

For US readers, the stakes are practical. American companies, schools, and government offices are already buying AI tools on the strength of vendor claims about which model is under the hood. A verifiable fingerprint would let a buyer, an auditor, or a regulator check that claim after the fact. It would also give smaller labs a way to prove their work without handing over their models. That does not settle the bigger debate about AI spending, where the [AI spending boom is already showing cracks](/article/ai-spending-boom-shows-cracks-as-big-tech-burns-cash) as big tech burns cash, but it does address one specific doubt: are you getting the model you paid for?

What happens next

Anthropic released a detailed technical specification and working code for various parts of the system, available under an open-source license on GitHub. Research is ongoing, and the company says there are no plans to deploy FlashProve at Anthropic or any other lab. The researchers are blunt about the status: "This research post is meant to share intermediate progress, and not a completed research agenda." They add: "We are sharing this explicitly with the goal that others look at what we have done, take it, and improve it."

The next test is whether other labs pick it up. A verification system only works if the labs whose models are being checked publish fingerprints and accept independent checks. Watch for other AI companies commenting on the specification, researchers stress-testing the cryptography, and any pilot where a vendor uses FlashProve to back a claim about its own model. Anthropic's own recent claim that [self-building AI is nearly within reach](/article/anthropic-says-self-building-ai-nearly-within-reach) shows why that kind of check may soon matter more, not less: as models do more consequential work on their own, proving which model did it becomes part of proving the work itself.