Anthropic builds system to prove which AI made an answer

Anthropic announces FlashProve system

Anthropic published a blog post Tuesday describing FlashProve, an open source system to verify which AI model produced a given output using cryptographic proof.

How FlashProve works

FlashProve provides provable inference, a cryptographic proof or mathematical guarantee that a given output came from a specific model, similar to email signatures confirming message origin.

Preventing model fraud

The system protects against impostors disguising one model as another, even when the impostor is the lab that built the model, preventing secret substitution of cheaper or expensive models.

Verifying third-party claims

FlashProve lets users verify claims by third parties that a specific task was done by an advanced model, such as vulnerability scanning or business plan review by appropriate models.

“A developer might claim that their software library was scanned for vulnerabilities by a frontier cybersecurity model, or a company might claim that its business plan was given the go-ahead by an appropriate model.”

Protecting proprietary model weights

The system requires labs to publish a fingerprint of each model that can be matched to outputs and verified cryptographically, without revealing the model's weights that define its behaviour.

Lightweight and open design

FlashProve is designed to be lightweight so verification can run on a local device. Anthropic released technical specifications and working code publicly on GitHub under an open source license.

Research in progress, not deployed

Anthropic has no plans to deploy FlashProve at its lab or any other lab. Researchers shared the work as intermediate progress and invited others to improve it.

“This research post is meant to share intermediate progress, and not a completed research agenda.”