Anthropic introduces FlashProve: a tool to verify which artificial intelligence actually responded

Anthropic introduced FlashProve, an open-source project aimed at addressing an increasingly important question: how to verify that a response was truly generated by the contracted artificial intelligence model? The initiative, announced in research published on November 2, proposes a mathematical verification of the origin of the responses.

The technology would allow detection if a provider silently replaces an advanced model with a cheaper one to cut costs. It would also help prevent results from being attributed to more powerful systems than those actually used.

A test to prevent deception in the artificial intelligence market

The researchers call the mechanism “provable inference”. Its operation is similar to cryptographic signatures that confirm the authenticity of an email: the user can verify if what they receive corresponds to the declared identity.

The difference is that FlashProve aims to verify operations of a model, even when the possible responsible party for a substitution is the very laboratory that manages it.

Anthropic provides concrete examples: a developer could claim that their software library was examined by an artificial intelligence specialized in cybersecurity, or a company could assert that an advanced model positively evaluated its business plan. Without an independent mechanism, clients must trust these statements.

How FlashProve works without revealing each company's secrets

To operate, the provider first publishes a cryptographic fingerprint associated with the internal parameters of a specific model. Then, when that system generates a response, it can provide verifiable proof that it was produced by the committed model.

The key is that the verification does not require revealing the so-called “weights”, the internal values that determine the system's behavior and constitute part of the intellectual property of the developers of closed models.

The design allows for commercial confidentiality to be maintained while simultaneously providing proof to clients, auditors, and third parties. The verification is intended to run on a conventional computer, without the need for specialized graphics hardware.

What results the prototype achieved and what its limits are

Anthropic published technical specifications and functional code on GitHub under an open license. In tests reported by the company, the system worked with Llama 4 Maverick, a model with 400 billion parameters, and produced demonstrations that can be verified using a CPU.

However, the technology still has a considerable cost: according to measurements from Anthropic, generating the proofs required between 70 and 95 times more GPU time than processing the same requests with a conventional tool. It also does not constitute a universal guarantee for any model or method of generating responses.

The project remains in the experimental phase. There is no implementation timeline or announced plans to integrate it into Claude or deploy it in other laboratories.

More transparency without sacrificing private innovation

The proposal addresses a specific commercial problem: when buyers cannot verify what technology they received, it becomes difficult to demand that the promised performance is met.

FlashProve could facilitate competition based on verifiable results, without forcing companies to disclose the internal workings of their models. It would also allow for the evaluation of third-party claims and the establishment of new auditing tools.

The authors clarified that this is an intermediate advancement and shared the work for other researchers to examine and refine. Its future utility will depend on technical evolution and whether providers adopt verification mechanisms.

For now, the main contribution is to demonstrate that transparency and the protection of intellectual property can advance together, even though it is still necessary to turn the prototype into a practical solution.