Anthropic is developing FlashProve, an open source system that aims to verify which AI model produced a given output, according to a blog post published Tuesday. The system is built to provide "provable inference," which the researchers describe as a cryptographic proof that an output came from a specific model.
According to the researchers, the system would let users confirm that an answer came from the model they think they are using. They say it would resist impostors, including the lab that built and runs the model. They note it could prevent an AI company from secretly swapping in a cheaper model to cut costs, or a more expensive model to inflate performance results. They also say it could let third parties verify claims that a task was done by an advanced model.
To work, AI labs would publish a "fingerprint" of each model they release. The researchers say these fingerprints do not reveal a model's weights, so closed-model makers could participate without disclosing their most valuable asset. The researchers say the system is lightweight enough to run verification on a local device.
Anthropic released a technical specification and working code under an open source license on GitHub. The researchers describe the work as intermediate progress, not a completed research agenda. They say research is ongoing and that there are no plans to deploy the system at Anthropic or any other lab. They wrote that they want others to "take it, and improve it."