Anthropic expands AI text watermarking across models

The invisible watermark, initially announced in August, is designed to identify text created by Anthropic’s Claude models. This rollout is a response to European Union regulations requiring transparency in generative AI tools.

Previously limited to newer models, the watermarking is now being integrated into older systems including Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6. The feature is mandatory for all users globally, regardless of location.

Currently, detection of the watermark is restricted to specific entities outlined in EU law, such as regulators, law enforcement, journalists, researchers, and companies needing to verify watermarking under the EU’s AI Act. General users cannot determine if their writing is flagged.

The initial announcement in August faced criticism, with concerns raised about the impact on tasks like proofreading and the potential for users to circumvent the system by rephrasing text with alternative AI tools.

Anthropic acknowledges limitations, such as the potential for shorter texts to lack sufficient watermark patterns. However, the company has refuted claims that the watermarking degrades output quality.

There are workarounds available for some users. By adjusting settings within the API, users of Opus 4.6 and Claude Sonnet 4.6 can remove the watermark. Additionally, users can select older models, including versions dating back to Opus 3 (released in 2023), which are not subject to the watermarking.

While other companies are also preparing to comply with EU regulations, Anthropic’s global implementation is notable. OpenAI, for example, has introduced textGrain to ChatGPT, but its automatic application is limited to Europe.