Anthropic develops open source system to verify AI model identity

Cryptographic proof for model outputs

The new tool provides a mathematical guarantee that a given answer came from a specific model. This mechanism functions similarly to cryptographic email signatures used to confirm message senders. It is built to stop impostors from disguising one model as another.

Protection against secret model swapping

Researchers say the system could prevent companies from secretly using cheaper models to cut costs. It also stops firms from inflating performance results by swapping in more expensive models. Third parties can use it to verify claims about specific tasks done by advanced models.

Requirement for published model fingerprints

The system works only if AI labs publish a fingerprint for each released model. These fingerprints allow later matching of outputs to their specific source. Verification remains possible even when the lab running the model attempts deception.