Anthropic shares FlashProve system to verify which AI made a reply

In short: Anthropic released an open-source research system called FlashProve to help people verify which AI model produced a specific response.

What happened

Anthropic published a research blog post describing FlashProve, a system meant to prove that a particular AI model, not an impostor, generated a given output. The goal is simple, if a tool says a reply came from a certain model, FlashProve aims to let anyone check that claim.

FlashProve uses cryptography, which is a way to use math to confirm something is authentic. Anthropic compares the idea to email signatures that help you confirm an email really came from the sender it claims to be from (like a tamper-proof seal on an envelope).

The system would require AI labs to publish a kind of public “fingerprint” for each model they release. Later, a user could match an AI output to that fingerprint and verify it, without needing access to the model’s private internals. Anthropic says these fingerprints do not reveal the model’s “weights”, meaning the private numbers inside the model that shape how it behaves.

Anthropic also released a technical specification and code on GitHub under an open-source license. The company says the research is ongoing, and it has no plans to deploy FlashProve at Anthropic or any other lab right now.

Why it matters

As more services rely on AI, it can be hard to know if you are getting what you paid for. A system like FlashProve could help catch cases where a company quietly swaps in a cheaper model to cut costs, or claims to use a stronger model to make results look better. It could also help customers verify third-party claims, like whether a security scan or business review was actually done by the model that was promised.

Source: TechCrunch AI