Anthropic is building a way to check which AI model actually produced the answer you’re reading, and it’s meant to work even if the company running the model is the one faking it. TechCrunch AI reports that the open source system, called FlashProve, was described in a blog post published Tuesday. It’s still research, and there are no plans to deploy it at Anthropic or any other lab. But the question behind it matters today: when you pay for a top model, how do you know that’s what you got?
Right now, you’re taking their word for it
When an API hands you an answer with a model name attached, that name is a claim. You usually can’t check it from the outside.
FlashProve goes after this with what researchers call “provable inference.” That’s a cryptographic proof, basically a math guarantee, that a specific model produced a given output. TechCrunch AI compares it to cryptographic email signatures, which let recipients confirm a message really came from its claimed sender.
The researchers point to a few scenarios:
- A lab secretly swaps in a cheaper model to cut costs.
- A lab swaps in a more expensive model to inflate its performance results.
- A third party says an advanced model did a job, and you have no way to check.
On that last one, the researchers write:
“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. How can a user verify these claims?”
What stands out here is who counts as a suspect. The design assumes the impostor could be the lab that built and runs the model. That’s an unusual thing for a frontier lab to plan around, and it’s the right call, because a check that only works if you trust the vendor doesn’t tell you much.
How it works
The process is simple at a high level:
- A lab publishes a “fingerprint” of each model it releases.
- Specific outputs can later be matched to that fingerprint and verified cryptographically.
- The check is designed to be light enough to run on a local device.
The smart part is what the fingerprint leaves out. It doesn’t reveal the model’s weights, the internal numbers that define how it behaves and that closed-model makers like Anthropic and OpenAI keep private. Think of a real fingerprint: it can identify you, but nobody can rebuild you from it. So in principle, commercial labs could take part without giving up their most valuable asset.
Alongside the post, the researchers released a detailed technical specification and working code for various parts of the system on GitHub, under an open source license.
What’s still missing
This isn’t a finished product, and the researchers say so: “This research post is meant to share intermediate progress, and not a completed research agenda.” A few gaps are worth keeping in mind:
- No numbers in the report. TechCrunch AI’s coverage doesn’t include benchmarks, such as what a proof costs to generate or how much it slows a model down. Those figures will decide whether this works at production scale.
- It needs labs to opt in. If a lab doesn’t publish fingerprints, there’s nothing to check against.
- Identity isn’t quality. A proof can tell you which model ran. It can’t tell you the answer was right, or that a vulnerability scan was any good.
What you can do with this
You can’t verify your model provider with FlashProve today, but it’s a useful prompt to tighten up how you work:
- Treat claims like “scanned by a frontier model” as marketing until there’s evidence behind them. Ask vendors for logs or other proof.
- If you build on an API or a reseller, run your own evals on a schedule. A sudden shift in behavior or quality is one of the few signals you’ve got right now.
- Add a provenance question to vendor reviews. Asking how a provider would prove which model served a request costs you nothing and shows you who has thought about it.
- If you work in cryptography or ML infrastructure, read the spec. The researchers are explicit about the goal: “that others look at what we have done, take it, and improve it.”
What to watch next
The real test is adoption. One lab publishing fingerprints wouldn’t get the industry very far, so watch whether any other lab picks this up, and whether anyone publishes hard performance numbers. My advice: don’t plan around FlashProve yet, but start asking vendors this question now. Buyers who demand proof are what will push labs to offer it. More details are available at TechCrunch AI.