Anthropic Is Developing Open-Source System FlashProve to Cryptographically Prove AI Output Provenance

November 4: According to IT之家 (ITHome), Anthropic is developing an open-source system called FlashProve that uses cryptographic proofs to verify which model generated a given AI output. The system aims to achieve "provable inference," giving users a mathematical guarantee that the answer they receive actually comes from the model they believe they are using, and defending against attempts to pass off one model as another.

FlashProve works similarly to cryptographic email signatures: it provides each model with a "fingerprint" that does not reveal its weights, which is then matched against a specific output for verification. The researchers say the system could prevent AI companies from secretly switching to cheaper models, or enterprises from switching to more expensive models to inflate test results, and could also verify third-party claims that a certain task was completed by an advanced model. The system is designed to be lightweight, allowing users to complete verification on local devices. Anthropic has published the technical specification and runnable code, which is now publicly available on GitHub under an open-source license. Research is still ongoing, and Anthropic has no plans to deploy the system yet.