Anthropic is building a way to prove which AI answered you

Anthropic says it is building a way to prove which AI model wrote a given answer. The system is called FlashProve, and it is meant to deliver what Anthropic calls "provable inference." In plain terms, users could confirm that a response came from a specific model. Anthropic described the work in a blog post on Tuesday, Nov. 3, according to PYMNTS.

The problem it targets

Anthropic's post says that when you talk to an AI model, you have almost no way to check who is answering. A user cannot confirm that a reply came from the model a developer claims is responding. A routing service could secretly swap in another model to cut costs. An attacker who tampers with the token stream, the flow of text coming back to you, could also change what you receive.

Anthropic adds that judging a model by its output alone is not reliable, even when the answers are good. Its approach uses cryptographic proofs, a kind of math-based receipt. A provider commits to a model, then proves that each response came from it.

What is not known yet

Anthropic says such proofs could support consumer products, internal security and checking compliance with AI pacing requirements. It also points to uses for auditors and regulators. The material does not say when FlashProve will be available, what it will cost, or which products will use it. For now, it is a system under development.