What jumped out at me here isn’t the crypto trick itself. It’s the mismatch between the promise and the reality of using it. Being able to prove which model answered a request sounds like exactly the kind of thing people want in regulated or adversarial settings, but if the verification step is expensive enough to make people hesitate, then we’re still stuck with the same old tradeoff: trust the system, or pay to verify it.
I’m interested because model identity is not a vanity feature. It matters when a vendor routes traffic between models, when you care about policy compliance, or when you need to know whether a response came from the model you actually paid for. That’s the practical pain point. If Anthropic really has a way to attach a cryptographic proof without exposing weights, that’s a meaningful technical step. But “meaningful” and “deployable” are not the same thing.
The cost caveat is doing a lot of work here. The source makes it sound like the proof mechanism is expensive enough that this won’t just become a default layer on every inference call. That’s the part I’d want to see stressed harder. A verification scheme that only makes sense for special cases is still useful, but it’s not the same as a general answer to provenance.
I also can’t help wondering how much of the appeal is psychological. A proof that says “this came from model X” is comforting, but it doesn’t tell you whether model X was the right model to use, whether the routing logic was honest, or whether the response was any good. It verifies identity, not quality. That distinction matters, and I think articles about model verification sometimes blur it.
If I were evaluating this for a real product, I’d ask one blunt question: how often do I actually need cryptographic certainty versus operational confidence? If the answer is “only for audits, disputes, or high-stakes workflows,” then expensive verification could still be a net win. If the answer is “for every request,” then this probably stays a research-flavored feature for a while.
Reference: Anthropic found a way to prove which AI model answered you — but there's a costly catch
