Anthropic shows a way to prove which AI model really answered you
Today you cannot check which AI wrote a reply. Anthropic's new free tool, FlashProve, adds a math proof to each answer. It is still early work.
When you talk to a friend in person, you can see and hear them. With an AI chatbot, you have almost no way to know what is really answering. On November 2, Anthropic shared early work on this problem.
Anthropic says a user today "cannot check that the responses came from the specific model that a developer claims is responding". A service could quietly swap in a cheaper AI to save money. An attacker could also change a reply on its way to you.
The fix Anthropic describes is a math proof that comes with each answer. First, the company that runs the AI locks in one model. Then it proves that each reply came from that model. Anthropic calls its new system FlashProve.
Anthropic says its own AI, Claude, helped build the system. In a test, one strong computer chip made proofs at around 110 tokens a second. Tokens are the small chunks of text an AI writes. Making the proof took about 70 to 95 times the chip time of writing the answer.
The proof does not show the model's private inner settings. Anthropic calls those settings the weights.
Anthropic shared the code for free and asked others to test it. It calls this "intermediate progress, and not a completed research agenda". Its goal is that others "take it, and improve it".
The post does not say that any chat app uses it yet. But it points to a future where you could check which AI really wrote your answer.
Open it and check the claim yourself. That is the point.