Anthropic has published an open specification for FlashProve, a system designed to prove which AI model produced a given output. The approach, described in a blog post on Tuesday, provides what researchers call provable inference: a cryptographic proof that an answer came from a specific model. For companies buying AI services, this matters because model identity directly affects quality, compliance and cost.
How FlashProve is supposed to work
The system works like cryptographic email signatures that confirm a message came from its claimed sender. Each output carries a proof that can be checked against a published fingerprint of the model. The check is designed to hold even when the party running the model has an incentive to misrepresent it. That property separates the proposal from ordinary logging or vendor assurances.
To enable verification, AI labs would publish a fingerprint for each released model. Later outputs can be matched to that fingerprint and checked cryptographically. The fingerprints do not disclose model weights, the internal values that define behavior and that closed-model developers keep private. Because of this, commercial labs could participate without giving away their most valuable asset.
The researchers describe two main misuse cases the design addresses. One is silent substitution by a provider: replacing the requested model with a cheaper one to cut costs, or with a more expensive one to inflate benchmark results. The other is unverified claims by third parties, such as a developer saying a software library was checked by a frontier cybersecurity model. Verification is designed to be lightweight enough to run on a local device.
What provable model identity means for buyers
For businesses that use AI in support, sales, legal review or security checks, model substitution is a procurement and risk issue. A cheaper substitute can lower answer quality and break process assumptions, while an undisclosed stronger model can distort evaluations and pricing. A verifiable link between output and model fingerprint would make contracts, audits and incident reviews more concrete. Small firms would gain a practical check without building their own evaluation team, while large firms could add it to vendor control.
The limits are substantial at this stage. Anthropic released a detailed technical specification and working code for various elements under an open-source license on GitHub, but research is ongoing and there are no plans to deploy it at Anthropic or another lab. Questions remain about adoption by other providers, performance overhead, integration with APIs and handling of modified or combined models. The announcement itself does not mean outputs available today can be verified this way.
The marker to watch is whether other labs publish model fingerprints and accept independent verification. A second signal is appearance of FlashProve-style proofs in procurement requirements, audit reports or regulated workflows. If those steps follow, model identity will shift from a vendor promise to a checkable fact.
