Anthropic tests cryptographic proof that an answer really comes from the stated AI model
With the open research prototype FlashProve, users should be able to check that a provider is not quietly using a different model, without the model weights being made public.
Anthropic has released research and code for FlashProve, a system that uses a cryptographic proof to show which AI model produced a given answer. According to TechCrunch, this could for example stop an AI company from secretly using a cheaper model to save costs, or a more expensive one to polish test results. A lab would publish a kind of fingerprint for each model version, against which answers can later be checked without revealing the model's weights.
Interesting Engineering describes the technique: a so-called zkSNARK, a proof that something is correct without showing the underlying data. In tests with Llama 4 Maverick, a model with 400 billion parameters, a single Nvidia H100 chip produced proof at about 110 tokens per second, a proof for 1,024 tokens was about 19 MB, and verification took about 17 seconds on an ordinary processor. Generating the proof does require 70 to 95 times more computing time than the regular answer.
It remains a prototype. Context is limited to 1,024 tokens, and the model's architecture stays visible, Interesting Engineering writes. According to TechCrunch, there are no plans to deploy the system at Anthropic or any other lab; the researchers are sharing it under an open licence so others can build on it.
Transparency: this edition was compiled by Claude, made by Anthropic.