Anthropic Unveils FlashProve to Verify AI Outputs, Releases Open-Source Code
Updated
Updated · TechCrunch · Nov 3
Anthropic Unveils FlashProve to Verify AI Outputs, Releases Open-Source Code
3 articles · Updated · TechCrunch · Nov 3
Summary
Anthropic said its FlashProve research system would let users cryptographically verify which AI model generated a specific output, aiming to prove model identity rather than trust a provider’s claim.
The design relies on labs publishing a model “fingerprint” that can be matched to outputs without exposing private weights, allowing verification to work even for closed commercial models.
Anthropic said the system is built to stop model-switching tricks, such as quietly substituting a cheaper model to cut costs or a pricier one to boost apparent performance.
Working code and a technical specification are already on GitHub under an open-source license, with verification designed to be lightweight enough for local devices.
The company said the work remains an intermediate research effort with no deployment plans yet, and is being shared so other labs and developers can improve it.