Can You Prove Which AI Model Answered?

1 A Target on a Roadmap

According to Anthropic's responsible-scaling roadmap, the company aims to have a provable-inference prototype by September 30, 2026. The date is a target, not a shipped capability: the roadmap describes a property the company is working toward, and any demonstration would be a prototype, not a production feature.

The concept is easy to state and hard to build: cryptographic proof that a specific output came from a specific model, without exposing the model's internals.

2 What Provenance Would Establish

Provenance answers an identity question. When a response, document, or decision surfaces in the world, verification would let a third party check which model generated it — distinguishing a genuine frontier-model output from a spoof, a smaller model masquerading as a larger one, or human-authored text presented as machine output.

That matters for accountability: if an agent takes an action with consequences, tying the action to a specific system is a precondition for assigning responsibility.

3 What It Would Not Establish

Verification of origin is not verification of content. A provenance proof confirms that model M produced output O; it says nothing about whether O is accurate, honest, or aligned with user intent. A model can be fully identified and fully wrong. Conflating the two is the central confusion.

4 How a Prototype Might Work

Approaches sketched in the AI-safety literature include cryptographic attestation of inference hardware, signed commitments to model weights, and zero-knowledge-style proofs that an output came from a committed model. Each carries a tradeoff: the stronger the proof, the heavier the computational cost and the narrower the class of outputs it can cover. According to the roadmap, the prototype's scope is deliberately limited.

5 The Hard Problem: False Before Signed

Provenance gets harder when the model itself is wrong. If a signed model hallucinated a claim, a valid proof would attach a trusted identity to untrusted content — making misinformation look more authoritative, not less. Wikipedia's summary of AI safety lists robustness and misuse alongside alignment — and a proof system that amplifies false outputs is a safety problem, not just an engineering one.

6 A Real Milestone With a Narrow Meaning

If the prototype lands on schedule, it would be a genuine infrastructure milestone: a building block for accountability in agent-driven systems. What it would not do is make model outputs trustworthy. Trustworthiness remains a property of the model's content, which provenance does not touch.

7 Reading the Roadmap Honestly

The disciplined reading is to treat September 30 as a checkpoint on a property that is proposed, not demonstrated, and to keep two questions separate: which model answered, and whether the answer was any good. Anthropic's roadmap, on its own terms, addresses only the first.