AI Output Verified: Anthropic's FlashProve Solution

Anthropic is developing an open-source system called FlashProve to cryptographically verify which AI model produced a given output. This technology aims to provide "provable inference," essentially a mathematical guarantee that an output originated from a specific model, akin to email signatures confirming sender identity.

The system is designed to be robust against malicious impersonation, ensuring users receive outputs from the intended AI model. It could prevent companies from secretly substituting cheaper models to reduce costs or employing more advanced models to artificially boost performance metrics.

FlashProve also facilitates the verification of claims made by third parties regarding the use of specific AI models for tasks like security scans or business plan approvals. For the system to function, AI labs must publish a model "fingerprint," which is a cryptographic identifier. Crucially, these fingerprints do not expose the model's private weights, allowing for compatibility with commercial models.

Researchers have released the technical specifications and working code for FlashProve under an open-source license on GitHub. While research is ongoing and the system is not yet deployed, Anthropic encourages external collaboration to improve the technology.