Anthropic is developing an open-source system called FlashProve designed to verify which specific AI model produced a given output. The system aims to provide "provable inference," using cryptographic proofs to offer a mathematical guarantee that an answer originated from a specific model. This mechanism would allow users to confirm the identity of a model and protect against impostors attempting to disguise one model as another.
The technology could prevent AI companies from secretly swapping models to cut costs or inflating performance results by using more expensive models. It also provides a way for third parties to verify claims regarding model usage, such as a developer claiming a software library was scanned by a frontier cybersecurity model.
To function, the system requires AI labs to publish a model "fingerprint" that can be matched to specific outputs. These fingerprints do not reveal a model's weights, allowing companies to protect their proprietary internal data while still enabling cryptographic verification. The system is designed to be lightweight enough to allow verification on a local device.
Anthropic researchers released a technical specification and working code for various elements of the system on GitHub under an open-source license. The company noted that the research is ongoing and there are currently no plans to deploy the system at Anthropic or other labs.
“This research post is meant to share intermediate progress, and not a completed research agenda,” the researchers wrote. “We are sharing this explicitly with the goal that others look at what we have done, take it, and improve it.”
