Anthropic Develops FlashProve to Verify the Identity of AI Models

Artificial intelligence company Anthropic is developing an open-source system designed to verify the identity of AI models behind generated responses.

The technology, called FlashProve, aims to make it harder for companies to misrepresent which AI models power their services. It uses cryptographic techniques to provide mathematical evidence linking an output to a specific model.

Researchers describe the approach as “provable inference”. It could give users and developers a way to check whether an answer actually came from the model a provider claimed to use.

FlashProve aims to prevent AI model substitution

Anthropic researchers say FlashProve addresses a growing transparency challenge in the AI industry.

Currently, users often have no independent way to confirm which model generated a response. This creates opportunities for providers to substitute models without informing customers.

For example, a company could replace a more advanced model with a cheaper alternative to reduce operating costs. Alternatively, a provider could claim that a premium model generated an answer to make its service appear more capable.

FlashProve is designed to make these substitutions easier to detect.

Reddit Tests Search Ads as AI Search Changes Online Traffic

The researchers also identify potential applications beyond consumer AI services.

Developers could use the technology to verify claims that software underwent security checks using a leading cybersecurity model. Similarly, businesses could provide evidence that a specific AI model reviewed their plans, reports or recommendations.

Cryptographic fingerprints protect proprietary models

The system would require AI developers to publish a cryptographic fingerprint for each model they release.

These fingerprints could then help verify whether a particular model generated an output.

Importantly, the process would not require companies to reveal their model weights. These numerical parameters shape an AI system’s behaviour and typically remain confidential at commercial AI companies.

This approach could allow companies such as Anthropic and OpenAI to participate without publicly exposing their proprietary model architectures.

Anthropic also says FlashProve is designed to be lightweight. This could allow users to verify outputs directly on their own devices rather than depending entirely on an external service.

Technical specifications and code released publicly

Anthropic has published a technical specification and working code for several FlashProve components.

The project is available on GitHub under an open-source licence. This allows researchers and developers to examine the technology, test its capabilities and contribute improvements.

However, FlashProve remains a research project rather than a commercially deployed product.

Anthropic has reportedly said it has no current plans to integrate the technology into its own services or deploy it at another AI laboratory.

The company sees the release as an opportunity for the wider research community to build on the work.

If developed further, technologies such as FlashProve could strengthen transparency and accountability across the AI industry. They could also help users distinguish between genuine model outputs and responses generated by substituted systems.