On 11 August Anthropic announced that all future Claude models will produce watermarked text. Google already applies its own method for Gemini, and OpenAI plans a similar approach.
On 11 August, Anthropic announced that all future Claude models will generate text carrying an embedded watermark identifying it as AI generated. Google already applies its own method to the output of its Gemini models, and Anthropic's approach builds on that design. OpenAI has not yet introduced a text watermark but plans to do so.
The move also responds to regulatory pressure: the European Union's AI Act requires watermarks on models released after 2 August 2026, covering images, audio, and video output as well. Image and video watermarks already achieve detection rates above 99 percent and have been used for years by OpenAI, Google, and Meta. Text is harder to mark, since a typical response spans only dozens or hundreds of words rather than the millions of pixels available in an image.
Whether this affects response quality remains disputed. Critics such as writer John Gruber call the watermark a distortion of the text, while researchers such as John Kirchenbauer of the Vector Institute argue that a detectable change is unavoidable as long as it does not noticeably reduce practical usefulness. Companies that rely on AI text in customer communication still need to determine whether watermarking changes how their content is perceived.
What this means for decision-makers
- Check whether the language models you use will soon embed text watermarks and how that affects internal approvals.
- Ask providers whether watermarking changes the quality of generated text in your specific use cases.
- Track implementation of the EU labeling requirement for models released after 2 August 2026.
This story was produced automatically from the source named above and checked by software before publication. The image is symbolic and shows neither the event nor a real person. How this paper is made