Research
Anthropic tests a way to check which AI is answering you, with no plans to use it yet
Anthropic has released FlashProve, an open prototype that uses math to prove which model gave an answer without revealing its secrets. It is slow and expensive, and no lab plans to use it for now.

The essentials
- Anthropic has openly released FlashProve, a research prototype that uses cryptography to prove which model generated an answer (MarkTechPost, TechCrunch).
- The lab publishes a fingerprint of the model that does not reveal its weights, the internal numbers that define how it behaves (MarkTechPost, TechCrunch).
- Claimed by one partyAccording to TechCrunch, neither Anthropic nor other labs have plans to use it: the research is still ongoing (TechCrunch).
Why it matters
Today, people who use an AI through a service have no way to check which model is answering them. A company could quietly swap in a cheaper one without telling anyone (MarkTechPost, TechCrunch). This points to a way to verify it, although for now it is slow and expensive.
The details
According to Anthropic's measurements, as reported by MarkTechPost, the system was tested with Llama 4 Maverick, a 400-billion-parameter model, on a single H100 chip (MarkTechPost). The catch is the cost: generating the proof takes roughly 70 to 95 times the chip time needed to answer without it. It also only handles short requests, of up to 1,024 tokens, the chunks the AI breaks text into (MarkTechPost).
What we don't know
- Whether any AI lab or service will adopt it.
- Whether experts outside Anthropic have reviewed its security.
- Whether it will work with long conversations and with other types of models.
Sources
- Anthropic Releases FlashProve: Cryptographic Proofs That a Reply Came From Committed Weights
- Anthropic is working on a system to prove AI models’ identity
Was this story useful? Yes Not really