Anthropic is developing an open source system called FlashProve to verify which AI model produced a given output. The project is meant to give users a cryptographic proof that an answer came from the specific model they think they were using, according to a blog post published Tuesday.
How FlashProve is meant to work
The researchers describe FlashProve as a form of provable inference. That means the system is intended to produce a mathematical guarantee that a given output came from a specific model. It is designed to work with commercial models, too, and to do so without forcing labs to reveal the model weights that define how the model behaves.
The system asks AI labs to publish a fingerprint for each model they release. Those fingerprints can later be matched to specific outputs and checked cryptographically. The fingerprints do not reveal the underlying weights, which closed-model makers such as Anthropic and OpenAI keep private.
What the checks are meant to stop
Anthropic says the system is built to resist impostors that try to make one model appear to be another, including cases where the impostor is the same lab that built and runs the model. The researchers say it could stop an AI company from quietly replacing a model with a cheaper one to reduce costs or with a more expensive one to improve reported performance.
The project also extends to claims made by third parties about work done by an advanced model. The researchers point to examples such as a developer saying software was scanned for vulnerabilities by a frontier cybersecurity model, or a company saying its business plan was approved by the right model. FlashProve is meant to let a user check those claims.
What Anthropic released
As part of the announcement, the researchers published a detailed technical specification and working code for parts of the system. The material is publicly available under an open source license on GitHub. The research is still ongoing, and there are no plans to deploy it at Anthropic or at any other lab.
The announcement also says the post is meant to share intermediate progress rather than a finished research agenda. The researchers say they are sharing it so others can review what they have built, use it, and improve it.
Licensed under CC BY-ND 4.0 International