Anthropic adds invisible watermark to Claude text as EU AI rules loom

Anthropic has begun embedding an invisible watermark in text generated by its Claude chatbot, a move designed to comply with new EU regulations. But the system has limitations and has yet to include a detection tool.

This article was drafted with AI assistance from multiple sources and was reviewed and approved by a human editor before publication.

Anthropic has started embedding an invisible watermark in text produced by its Claude chatbot, a step the company says is meant to comply with new European Union regulations. The watermark is designed to be imperceptible to humans and to survive copy-pasting, but it does not necessarily prove that Claude authored the text from scratch. "Claude may not be the original author. People often use Claude to proofread, translate, summarize, or convert files. The output can carry a Claude mark even if the underlying ideas, text, or data originated from another source," the company said in a statement.

The watermark is applied to all Claude models released on or after August 2, and older models will be upgraded to include it. It is being rolled out worldwide. For supported file types such as .jpg, Claude also attaches signed provenance metadata using an open standard.

Anthropic's move comes as the EU introduces regulations that require machine-readable markings on AI-generated content. Violations carry the risk of stiff fines. Late last month, Anthropic was among roughly 190 organizations that signed a voluntary EU code of good practice, alongside Google, Meta, Microsoft, Mistral, and OpenAI.

However, the watermark has notable limitations. It may be absent if the text is very short or heavily edited, and it can be rendered undetectable by extensive editing, reformulation, translation, or mixing with other text. Proofreading or translating with Claude can also leave watermark traces, which complicates the picture further.

Anthropic has not yet released a watermarking detection tool, even though the EU code calls for one. Google, by contrast, has developed SynthID, a nonpublic advanced digital watermarking and detection technology, and is pushing Apple, Nvidia, and OpenAI to adopt it as a standard.

The watermark arrives amid growing concerns about AI-generated content fraud. A survey by the BoD found that 52% of authors now use AI tools for writing projects, up from about a third in 2024, and that daily AI users among authors have more than doubled to 23%. In one prominent case, Hachette Book Group stopped selling the horror novel 'Shy Girl' over AI allegations, calling it a contract breach. AI detector Pangram estimates that up to 78% of the book could be AI-generated, while author Mia Ballard denies the allegations, saying a freelance editor inserted AI material without her knowledge.

In education, the impact is also being felt. A Hochschule Darmstadt survey found that 92% of students in Germany use AI tools at least occasionally, and 65% worry that AI could weaken their cognitive abilities. A Center for Universal Education study warns that AI models can undermine learning if children outsource their thinking. In response, Danish students aged 16-19 must now orally defend their written work, and the University of Chicago Law School has banned laptops for first-semester students.