Full story
EU watermarking begins Aug 2
Anthropic says future Claude models will embed a hidden watermark in text generated in the European Union, starting with Claude models available in the EU since August 2.
“undetectable to the reader, but is detectable to anyone who has a key that encodes it”
In its explanation of how the system works, Anthropic describes “low-stakes choices” where Claude can choose between words like “overcast” and “grey,” creating a pattern that is “undetectable to the reader, but is detectable to anyone who has a key that encodes it.”
Anthropic also says the watermark is integrated into the content so it “will travel with the text when it is copied and pasted and can persist even with some editing,” including when users ask Claude to edit, translate, summarize, or transform material from another source.
The company adds that watermarking does not change quality, stating “Watermarking does not impact the quality of Claude’s output,” and it will not add hidden characters or require extra tokens.
Anthropic further says it is building a detection API and that other major model developers have signed the same Code of Practice and will implement their own watermarks.
Editing, translation, and code
Anthropic warns that the watermark can appear even when Claude is only used to edit someone else’s writing, because “Claude may not be the original author. People often use Claude to edit, translate, summarize, or convert files.”
TechCrunch reports that Anthropic told users it is possible to hide the watermark with rewriting, but that “light editing probably won’t remove the watermark completely,” while “a complete rewrite where every word is replaced will.”
On detection limits, Anthropic acknowledges that identification is more complex when a text undergoes significant edits, including “Extensive edits, translations, very short fragments, or the incorporation of content within a longer text can affect detection.”
For code specifically, Anthropic says code should have less watermarking than other text because the model must produce working code, but it can still apply the watermark in areas with arbitrary word choice such as “comments within code,” where it will have “a negligible effect on the actual code produced.”
TechCrunch also notes that Anthropic frames watermarking as different from AI text detectors that look for “tells” in writing, emphasizing that “Picking up on these patterns is fundamentally different from checking for a watermark.”
User backlash and compliance stakes
As Anthropic rolled out watermarking in the EU since August 2, TechCrunch reported that the move sparked debate on Reddit and that Business Insider said “dozens” of users on X claimed to cancel their Claude subscriptions.
“uncomfortable" with Anthropic's new watermarking plan”
Business Insider quotes users describing practical concerns about authorship and client work, including Arturo Villarroya, who said he canceled his Claude Max subscription after becoming “uncomfortable” with Anthropic’s new watermarking plan.
In Business Insider’s reporting, Villarroya worried that a watermark could remain even when the underlying writing was original and Claude was used only to translate, check spelling, fix academic citations, or lightly summarize, and he said its presence could create problems in academic and professional settings with strict rules on AI use.
Anthropic’s own blog explanation counters that the watermark indicates Claude processed text rather than necessarily who created the original work, stating “Anthropic said marks can remain after Claude proofreads, translates, summarizes, or otherwise edits content.”
The stakes, as TechCrunch frames them, are tied to the EU AI Act’s Transparency Code, which requires systems that make it possible to identify AI-generated content, and Anthropic says “other major model developers have signed the same Code of Practice and will be implementing their own watermarks.”
