.png){kind=logo}
Anthropic is working on an open-source system that could help users verify which AI model generated a particular response. The system, called FlashProve, uses cryptographic technology to prove that an output came from a specific AI model, addressing concerns about transparency and trust in AI services.
The tech giant announced it in a research blog post on Tuesday, November 3, 2026. The technology lets users independently check whether the model they interacted with was the one the provider claimed.
FlashProve is built around what researchers call ‘provable inference.’ It uses a cryptographic proof, essentially a mathematical guarantee, to establish that a specific AI model generated a particular response.
The approach is similar to cryptographic email signatures that allow recipients to verify whether a message genuinely came from its claimed sender.
Researchers explained that the system could prevent AI providers from secretly replacing a model with another version. For example, it could prohibit a company from substituting a cheaper model to reduce operating costs or using a more expensive model to inflate performance claims.
The technology is also intended to help verify third-party claims about AI-assisted work.
Also Read: Anthropic’s Quarterly Loss Shrinks as Revenue Surges Ahead of IPO
For FlashProve to work, AI companies will need to publish a unique ‘fingerprint’ for every model they release. These fingerprints can then be matched against specific outputs and verified through cryptographic methods.
Additionally, the process does not require companies to reveal their model weights, the internal numerical parameters that determine how an AI model behaves. Companies such as Anthropic and OpenAI keep these weights private as they are among their most valuable assets.
This could allow the system to work with commercial AI models without requiring developers to disclose proprietary technology. Anthropic also designed FlashProve to be lightweight enough for verification to run on a local device.
Alongside the announcement, Anthropic's researchers released a detailed technical specification and working code for different parts of FlashProve. The material is publicly available under an open-source license on GitHub.
However, the company clarified that the project remains a research effort. There are currently no plans to deploy FlashProve at Anthropic or any other AI lab.
The researchers said the announcement was intended to share intermediate progress rather than present a completed research agenda. They also encouraged other researchers and developers to examine the work, use it, and improve the system.
If developed further, FlashProve could offer businesses and individual users another way to assess the reliability of AI services, particularly when providers claim their models power their products.