Anthropic has published details about an open-source verification system designed to confirm which AI model generated any particular output. The system, named FlashProve, aims to deliver what its researchers term "provable inference"—essentially a cryptographic proof, or mathematical guarantee, that a specific model produced a given result. The goal is to allow users to confirm that the response they received genuinely originated from the model they believe they are interacting with.

Much like cryptographic email signatures that allow recipients to verify a message truly came from its stated sender, FlashProve would resist impostors attempting to pass one model off as another—even when that impostor is the very lab that built and operates the model. The researchers specifically point out that the system could prevent an AI company from covertly substituting a cheaper model to reduce costs, or a more expensive one to boost its performance benchmarks. FlashProve could additionally be used to validate third-party claims that a particular task was handled by an advanced model. As the researchers write, "A developer might claim that their software library was scanned for vulnerabilities by a frontier cybersecurity model, or a company might claim that its business plan was given the go-ahead by an appropriate model. How can a user verify these claims."

For the system to function, AI labs must publish a "fingerprint" for each model they release, which can later be matched against specific outputs and verified cryptographically. Importantly, these fingerprints do not expose a model's weights—the internal numerical parameters that govern its behavior and that closed-model developers like Anthropic and OpenAI keep confidential. This means labs could publish fingerprints without surrendering their most valuable intellectual property, making the system viable for commercial models as well. The system is also engineered to be lightweight, allowing verification to run on a local device.

Alongside the announcement, the researchers released a detailed technical specification and functional code for various components of the system, made publicly available under an open-source license on GitHub. Work on the system continues, and there are currently no plans to deploy it at Anthropic or any other lab. "This research post is meant to share intermediate progress, and not a completed research agenda," the researchers note in the announcement. "We are sharing this explicitly with the goal that others look at what we have done, take it, and improve it."