Anthropic is developing FlashProve, an open-source system intended to verify which AI model produced a particular response. In a blog post published Tuesday, the researchers described it as a way to provide “provable inference”: a cryptographic proof that links an output to a specific model.
The system is designed to make it harder to pass off one model’s responses as another’s, including when the substitution is made by the company operating the model. The researchers say it could help users detect cases such as a provider secretly switching to a cheaper model, or using a more expensive one to improve reported performance.
FlashProve could also help check claims that a particular model was used for a task. Anthropic’s researchers give examples including a software library said to have been reviewed for security vulnerabilities by an advanced model, or a business plan said to have been assessed by an appropriate model.
For verification to work, AI labs would publish a fingerprint for each model they release. The fingerprint could be matched to outputs and checked cryptographically, without disclosing the model’s weights—the internal values that shape its behavior. The researchers say the system is intended to be lightweight enough for checks to run on a local device.
Anthropic has released a technical specification and working code for parts of FlashProve on GitHub under an open-source license. The research is still in progress, and the company says it has no plans to deploy the system at Anthropic or elsewhere. The researchers presented the release as intermediate work and invited others to examine and improve it.
Sources: TechCrunch.