Anthropic has introduced FlashProve, a research project for an open system designed to verify the origin of artificial intelligence responses. Using cryptographic proof, users will be able to confirm that a result was actually generated by the model they were promised, rather than a cheaper or less powerful substitute.

The project aims to address the issue of trust in AI services: today, users cannot always verify exactly which model processed their request. This is especially important for companies that use AI for critical tasks or claim to have obtained a result from a cutting-edge model, TechCrunch reports.

How to prove which model answered a request

FlashProve implements an approach researchers call provable inference. Essentially, it provides cryptographic confirmation that a particular response was generated using a specific model.

Anthropic compares the idea to digital signatures for email. These allow recipients to verify that a message was actually sent by its purported sender. Similarly, FlashProve is intended to confirm the origin of an AI response and protect against attempts to pass one model off as another.

This verification could be useful even when the company that developed and operates the model is itself attempting the substitution. For example, an AI service provider could theoretically replace an expensive model with a cheaper one without users noticing, cutting costs while maintaining the same promises to users. The reverse situation is also possible: a company might claim results from a more powerful model while actually using a different one.

FlashProve is designed to make such substitutions detectable.

Why this matters beyond AI services

The system could be useful for more than just verifying chatbot responses. It would also make it possible to verify third-party companies' claims about which models were used to perform particular tasks.

For example, a software developer might claim that a cutting-edge cybersecurity model checked their library for vulnerabilities. Or a company might say that its business plan received approval from an AI system suited to that type of assessment.

In both cases, the same question arises: how can anyone confirm that the stated model was actually involved in the work?

Anthropic's researchers believe that cryptographic verification of a result's origin could give users that ability. This is especially important as AI solutions begin to be used in processes where trust in the quality and origin of a response has practical implications.

A model fingerprint without revealing its inner workings

For FlashProve to work, AI developers will need to publish a special digital “fingerprint” for each model they release. This fingerprint can then be compared against the system's outputs and verified cryptographically.

The fingerprint does not reveal the model's weights—the internal numerical parameters that determine its behavior. Developers of closed systems, including Anthropic and OpenAI, generally do not publish this data because it is one of their most valuable commercial assets.

Companies will therefore be able to participate in the system for verifying the origin of responses without providing access to their models' inner workings. This is an essential condition for applying the technology not only to open AI products but also to commercial ones.

Another requirement for FlashProve is low computational overhead. Proof verification must run directly on the user's local device, without needing to access the model developer's powerful infrastructure.

Anthropic has released the code but has no plans yet to deploy the system

Alongside the announcement, the researchers published a detailed technical specification for FlashProve and working code for individual project components. The materials are publicly available on GitHub under an open-source license.

However, the project remains at the research stage. Anthropic has not announced plans to integrate FlashProve into its own products or make the system available for practical use by other AI companies.

The authors emphasize that the published materials represent an intermediate stage of the work, rather than a completed research program. They hope other specialists will examine the proposed approach, test its capabilities, and suggest improvements.

If FlashProve can be developed into a practical tool, it could become an important means of bringing transparency to the AI market. Today, companies compete primarily on the quality and capabilities of their models. But as these services become more widespread, another question will become increasingly important: did the user actually receive a response from the system they paid for or were promised?