FlashProve: Open-source prototype verifies the origin of AI-generated responses
Anthropic has launched FlashProve, an open-source tool designed to identify which artificial intelligence model generated a specific response. The aim is to enable users and businesses to confirm whether the result they receive matches the contracted system, even when the service involves intermediaries. The details were disclosed in a company statement.
The solution can detect model substitutions and changes in responses. For corporate clients, this means being able to verify whether an analysis was performed by the chosen AI. For independent auditors, the tool helps ensure that the system being tested is the same one offered to end users, according to Anthropic.
How verification works
The process begins with the provider publishing a cryptographic commitment—a kind of fingerprint linked to the exact set of model weights. These weights are the numerical parameters that define the AI’s behavior and remain protected.
For each generated response, the provider creates a proof that demonstrates the correspondence between the result and the registered model. The recipient of this proof can validate it on a standard CPU without needing access to the weights or the provider’s processing infrastructure. The mechanism was created to prevent a response from another model from being accepted as authentic.
Cost remains a barrier
The technical report shows that verification still requires significant resources. According to researchers, generating the proofs consumed approximately 70 to 95 times more GPU time than was needed to process the same inputs in the adopted reference.
The proofs were about 19 megabytes in size. Verification took 17 seconds with 24 CPU threads and 21 seconds with a single thread. Generation occurs after the response is produced, in a separate process, which reduces interference with user service but allows a tampered response to arrive before the problem is detected. Thus, the system functions as an additional verification layer, not as an immediate block against manipulations.
Current project status
The official repository classifies the code as a research artifact, with no maintenance or receipt of external contributions. The test uses Meta’s Llama 4 Maverick model and a context window of 1,024 tokens.
Anthropic acknowledges that the prototype still has limitations and presents preliminary results. The certification only verifies the correspondence between the model and the response, without attesting to the veracity of the generated content. For businesses, the advancement lies in the possibility of auditing the origin of the AI service they use.
Frequently Asked Questions
What is FlashProve?
It is an open-source prototype that allows verification of which AI model generated a response, without exposing its internal weights.
Does the system guarantee that the response is correct?
No. It verifies the correspondence between the response and the model, but does not attest to the veracity of the produced content.
Is FlashProve already a commercial product?
No. The repository classifies it as a research artifact, with no maintenance or receipt of contributions.








