According to a blog post published Tuesday, Anthropic is working on an open-source system to verify which AI models produced a particular output. The system, called FlashProve, is built to provide what researchers call “provable inference,” a cryptographic proof, essentially a mathematical guarantee, that a particular output came from a particular model. This is intended to help users confirm that the answers they receive really come from the model they think they are using.
Similar to cryptographic email signatures that allow recipients to verify that a message is actually from the sender it claims to be, FlashProve is robust against impostors attempting to pass off one model as another, even if the imposter is the lab that built and ran the model. In particular, the researchers note that the system could be used to prevent AI companies from secretly switching to cheaper models to cut costs or switching to more expensive models to inflate performance results.
FlashProve can also be used to verify claims by third parties that a particular task was performed by an advanced model. “Developers may claim that their software libraries were scanned for vulnerabilities by state-of-the-art cybersecurity models, and companies may claim that the right models gave their business plans the green light,” the researchers wrote. “How can users verify these claims?”
For this system to work, the AI lab must publish a “fingerprint” of each model it releases, which can later be cryptographically verified against specific outputs. Importantly, fingerprinting does not reveal model weights, internal numbers that define model behavior, or information that closed model makers like Anthropic and OpenAI keep private. This means that labs can publish their fingerprints without giving up their most valuable assets, and the system can also work on commercial models. The system is designed to be lightweight, allowing you to perform validation on your local device.
As part of the announcement, the researchers published detailed technical specifications and working code for various elements of the system, which are available on GitHub under an open source license. Research on this system is ongoing, but there are no plans to introduce it to Anthropic or any other lab.
“This research post is intended to share interim progress and not a completed research project,” the researchers wrote in their announcement post. “We clearly share this with the goal that others will see what we have done, accept it, and improve.”
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
