Can you prove your AI answer came from the model you paid for? Anthropic's new research says yes

When you use an AI model today, you have almost no way to confirm which model actually produced the answer. According to Anthropic, a response could come from a cheaper model quietly substituted by a routing service, or from a stream of text tampered with by an attacker. Judging the output alone is unreliable, even when the answers look high quality.

Anthropic and researchers from EPFL, MIT and other institutions describe progress on cryptographic proofs of inference. These let a model provider commit to a specific model and then prove that each response was produced by it. The team focuses on a type of proof called a zkSNARK, which they say suits this job because of its strong hiding and efficiency properties. Their main system is called FlashProve. The authors say such proofs could support uses from consumer products to internal security, and could help verify compliance with AI pacing requirements.

The work is described as intermediate progress, not a finished research agenda. Anthropic has released its code, Lean proofs and documentation under the Apache 2.0 licence, and says it hopes others will examine, reuse and improve the work. For now this is research rather than a product, but it points toward a future where businesses can check what AI service they are really getting.