Artificial intelligence company Anthropic is developing an open-source system designed to verify the identity of AI models behind generated responses, potentially making it harder for companies to misrepresent which models are powering their services.
The system, called FlashProve, is based on what researchers describe as “provable inference”. It uses cryptographic techniques to provide mathematical evidence that a particular output was generated by a specific AI model.
The aim is to allow users and developers to check whether an answer actually came from the model they were told was being used.
How FlashProve could prevent AI model substitution
Anthropic researchers say FlashProve could help address a transparency problem in the AI industry: users often have no independent way to confirm which model generated a response.
The system is designed to resist attempts to pass off one model’s output as that of another, including cases in which the deception could come from the company operating the models.
For example, an AI provider could potentially replace a more capable model with a cheaper alternative to reduce operating costs without informing customers. Alternatively, a company could claim that a more expensive or advanced model produced results to make its performance appear stronger.
FlashProve is intended to make such substitutions easier to detect.
The researchers also see potential applications beyond AI services. Developers could use it to verify claims that software was checked for security vulnerabilities by a leading cybersecurity model, while businesses could provide evidence that a particular AI model reviewed their plans or recommendations.
Model fingerprints keep proprietary weights private
For the system to work, AI developers would need to publish a cryptographic fingerprint for each model they release.
These fingerprints could later be matched against outputs to verify their origin. Importantly, they would not disclose the model’s weights, the internal numerical parameters that shape its behaviour and are typically kept confidential by commercial AI companies.
This could allow companies such as Anthropic and OpenAI to participate without making their proprietary model architecture publicly available.
Anthropic also says the system is designed to be lightweight, allowing verification to take place on a user’s own device rather than relying entirely on an external service.
Technical specifications and code released publicly
Alongside the announcement, Anthropic published a technical specification and working code for several components of FlashProve.
The project is available on GitHub under an open-source licence, allowing other researchers and developers to examine, test and improve the system.
However, FlashProve remains a research project. Anthropic said it has no current plans to deploy the technology within its own services or at another AI laboratory.
The company described the release as an opportunity for the wider research community to build on its work, stressing that the project represents ongoing progress rather than a completed research programme.
