What is FlashProve? Anthropic’s open-source tool to prove which AI model generated a response

Anthropic has announced FlashProve, an open-source system for ‘provable inference’ that works like a cryptographic email signature for AI outputs. It aims to let users verify that a response came from the model a company claims to have, without exposing the model’s weights. 

Anthropic is developing an open-source system, FlashProve, to help users verify which artificial intelligence (AI) model generated a particular response. Announced in a research blog post on Tuesday, November 3, FlashProve is designed to verify the authenticity of AI-generated outputs using cryptographic proofs, enabling users to determine whether an answer is truly from the intended AI.

The system, dubbed ‘provable inference’, is intended to prevent AI companies and third parties from presenting incorrect interpretations of the models behind their services. FlashProve could help mitigate fears that companies are covertly substituting high-cost, high-performance AI models with lower-cost versions or falsely touting the capabilities of the models being used in certain applications, Anthropic researchers said.

Anthropic's FlashProve aims to prevent AI model impersonation

FlashProve operates on a principle similar to cryptographic email signatures, which enable a recipient to verify the sender of an email. The system is working to provide greater security from the passing off of one AI model as another, including by the company running the AI model.

The technology could also be used by users to verify claims from businesses and developers. For instance, a software company could prove that its code was examined by a particular advanced cybersecurity model. Likewise, companies might show that a particular AI system took into account their plans or suggestions.

How FlashProve verifies AI-generated responses

Another key feature of FlashProve is that it is designed to be lightweight, enabling verification on the user's device without heavy reliance on external infrastructure.

Pragya Singha Roy