Anthropic adds text watermarking to future Claude models

Anthropic has begun adding a watermark to text generated by future Claude models, according to the company. The system uses a cryptographic key and preceding text to create a statistical pattern that a party with the matching key can test for. Anthropic says the watermark is invisible to readers and adds no tokens. The company tied the rollout to the EU transparency code it signed in July and said it is deploying globally because it cannot yet limit the feature to the EU. Anthropic also plans an API for checking text for the signal.