Catch me if you chatbot

Anthropic is working on a new open-source tool that could help people check whether an AI response really came from the model they were promised.

Called FlashProve, the tool uses mathematical proof to confirm which AI model produced an answer.

This could stop companies from secretly using cheaper models to save money or switching to more powerful ones to make their results look better.

Think of it as a digital signature that confirms an AI response came from the right model.

Companies wouldn't have to share how their AI models work.
The tool is designed to let users verify responses on their own devices.
Anthropic has shared the code and technical details, but there are no plans to launch it yet.

AI companies would publish a unique digital fingerprint for each model.

This would allow users to check where an answer came from without exposing the private technology behind it.

FlashProve could also help businesses check whether a specific AI model was used to review software for security problems or assess business plans.

Anthropic hopes other developers will build on the research and help improve the tool.

I’ll believe it when I see it. -MV