Anthropic's FlashProve: math proof of which model really answered
When you use an AI through an app, how do you know which model really answered? You don't. A company could swap in a cheaper model, or someone could change the reply on the way to you.
Anthropic published research on November 2 on a system called FlashProve, with collaborators from EPFL, MIT and others. It uses zkSNARKs, a kind of cryptographic proof that shows something is true without showing the secret behind it. Here, it proves a reply came from a specific set of model weights, without revealing those weights. It was tested on a Llama model.
TechCrunch says Anthropic is 'working on' it, so this is research, not a product you can use today. Bon, it is a bit like checking that your maple syrup really comes from Québec, but with math.
Independent checks are exactly what AI needs. With proofs like this, evaluators and users would not have to take a lab's word for it. Good news, and still: a proof tells you which model answered, not whether that model is safe.
