Anthropic described the system in a blog post published on Tuesday, TechCrunch reports.

FlashProve is built to offer what researchers call “provable inference,” TechCrunch says. That is a cryptographic proof that a given output came from a specific model. It would let users confirm that an answer really came from the model they think they are using.

The system works a bit like digital email signatures, which let people check that a message came from its claimed sender, the site reports. It is meant to resist anyone posing one model as another, even the lab that built and runs it.

The researchers say it could stop an AI company from secretly swapping in a cheaper model to save money, TechCrunch reports. It could also stop a company from using a pricier model to puff up its test results.

The system could also check claims that an advanced model did a certain task, the site says.

To work, AI labs would publish a fingerprint of each model they release, TechCrunch reports. The fingerprint could later be matched to outputs. It would not reveal a model's weights, the private numbers that shape how it behaves.

The system is also meant to be light enough to run checks on a local device, the site says. The researchers released a technical plan and working code on GitHub under an open source license.

Research is ongoing, and there are no plans to deploy it at Anthropic or any other lab, TechCrunch reports. The researchers said they were sharing early progress so others could take it and improve it.

Only TechCrunch has reported this so far, and NewsHabit hasn't seen independent confirmation.

What we know: Anthropic researchers shared early work and code for a system called FlashProve.
What we don't: Whether any AI lab will adopt it. No second outlet has confirmed the details yet.