Anthropic is developing a new technology called FlashProve designed to verify if responses were genuinely produced by the artificial intelligence model a company claims to use. The system employs cryptography to generate a mathematical proof of the model's identity behind specific outputs.

The primary goal is to give users a way to confirm they are interacting with the AI announced by the vendor instead of a different or less advanced version. According to Anthropic, this mechanism could also make it harder for companies to secretly substitute their models with cheaper alternatives or present performance results that do not match the technology actually used.

FlashProve utilizes the concept of verifiable inference, which involves generating a cryptographic proof that a specific response originated from a particular model. This functions similarly to digital signatures used in emails to confirm the sender's identity, but here it identifies which AI produced content. To achieve this, model developers would need to publish a kind of fingerprint for each released system, which can then be compared against AI outputs using cryptographic mechanisms.

A key feature of this technology is that these fingerprints will not reveal the model weights—the internal values determining its functioning and behavior. Companies like Anthropic and OpenAI typically keep such information private but could participate in a verification system without compromising trade secrets. Additionally, the technology is designed to be lightweight, allowing verification to occur directly on a local device.

One significant problem FlashProve intends to address is the possibility of a vendor presenting one model to clients while utilizing another during service operation. For example, a company might advertise access to an advanced AI but switch to a cheaper alternative to reduce costs. The mechanism could also hinder situations where a firm uses a different model than claimed when demonstrating performance.

Verification capabilities extend beyond services offered directly by AI companies. Researchers cite examples such as a developer claiming to have used an advanced cybersecurity model to scan software libraries for vulnerabilities, or a business declaring that its business plan received approval from an appropriate AI. With FlashProve, it would be possible to cryptographically verify if the announced model was actually used in these processes.

Despite the public presentation, FlashProve remains a research project. Anthropic has made available a detailed technical specification and functional code for various components under an open-source license, enabling the community to examine and develop the system. However, the company has not announced plans to implement the technology in its own services or other AI platforms. Researchers state that the publication represents an intermediate advancement in their work rather than the conclusion of a research agenda.