Anthropic's FlashProve aims to prove which AI model wrote an output

[Python35]

How do you know which AI model produced an answer? TechCrunch reports that Anthropic is working on FlashProve, an open source provable inference system that gives a cryptographic proof of which model generated a given output

Under the system, AI labs would publish a fingerprint for each model, and outputs could be checked against it. The fingerprint does not reveal the model's weights, so closed models can take part, and verification is light enough to run on a local device

Possible uses include catching providers that quietly swap in a cheaper model, and checking claims such as a particular frontier model having audited a software library. The system is designed to resist impostors, including the lab that runs the model

The specification and working code have been published on GitHub. Anthropic described the research post as sharing intermediate progress rather than a completed research agenda, and there are no deployment plans yet at Anthropic or any other lab

As AI is used in more important decisions, being able to verify which model produced an output could become very valuable. It would help businesses check they are getting what they pay for. Making it open source gives other labs a chance to adopt it too

Would provable model identity build trust in AI? And should AI providers be required to prove which model they use?

Anthropic is working on a system to prove AI models' identity | TechCrunchtechcrunch.com