Anthropic is building a cryptographic proof of which AI model wrote an answer

Anthropic detailed the project, called FlashProve, in a blog post published Tuesday. The system aims to deliver what the researchers call “provable inference” — a mathematical guarantee that a specific output came from a specific model, so a user could confirm the answer they got really came from the model they think answered it. The researchers compare it to a cryptographic email signature: a way for whoever receives an AI output to confirm its origin, resistant even to an AI lab trying to disguise one of its own models as another.

The researchers point to several practical uses. FlashProve could stop an AI company from quietly substituting a cheaper model to cut costs, or a pricier one to flatter its benchmark results, without customers finding out. It could also let outside parties verify claims about which model performed a task — the blog post gives the example of a developer claiming a code library was scanned by a frontier cybersecurity model, or a company claiming its business plan was reviewed by a particular model, claims a user currently has no way to check.

Technically, the system requires AI labs to publish a cryptographic “fingerprint” for each model they release, one that can later be matched against specific outputs. Crucially, those fingerprints wouldn’t expose a model’s weights — the internal parameters that make closed developers like Anthropic and OpenAI treat them as trade secrets — so labs could adopt the scheme without giving up what they consider their most valuable asset. The design is also meant to be lightweight enough to verify on a local device.

Anthropic released a detailed technical specification and working code for parts of the system on GitHub under an open-source license. The researchers describe the work as unfinished: “This research post is meant to share intermediate progress, and not a completed research agenda,” they wrote, adding that they want others to build on it. Anthropic said it has no current plans to deploy FlashProve itself, and the post doesn’t name any other lab committing to use it either.

Why it matters

If labs eventually adopt something like FlashProve, it would give enterprises and regulators a way to audit AI-model claims that today rest entirely on trust in the vendor — a gap that matters more as companies increasingly build security, compliance and financial workflows on top of specific models’ claimed capabilities.