Anthropic details how Claude text watermarks work
Anthropic says Claude’s hidden text watermark is meant to meet EU AI Act rules, and that light editing may weaken it without fully removing it.
Anthropic has shared new details on how Claude will watermark its text output after saying earlier this week that it would adopt the measure to comply with the EU AI Act’s Transparency Code. The update has already stirred debate among Claude users, because it raises a practical question: how easy will AI-generated text be to identify, and how much control will people still have over what they publish?
The company’s latest explanation focuses on how the watermark is supposed to work, how it could be detected, and where it stops being effective. It also addresses worries about whether the system could affect code generation — an important point for people who use Claude for programming as well as everyday writing. According to IT-PUB News, Anthropic presented the change as part of its effort to meet the EU transparency rules.
Anthropic says the watermark will stay invisible to readers
In a blog post on Friday, Anthropic described watermarking as a way to build a hidden pattern into Claude’s responses when the model makes “low-stakes choices” between words or phrases. The company used the example of choosing between “overcast” and “grey” to describe the weather.
Anthropic said this pattern would be “undetectable to the reader, but is detectable to anyone who has a key that encodes it.” The company also said the watermark would not change the quality of Claude’s output.
“To a reader, a watermarked response is indistinguishable from an unwatermarked one,” Anthropic said.
That matters because Anthropic is not talking about a visible label or a simple disclaimer. The idea instead is to make AI-generated text identifiable through a technical system working in the background.
EU AI Act rules are driving the change
Anthropic said the watermarking is being introduced to comply with the EU AI Act’s Transparency Code, which requires AI companies to use systems that make it possible to identify AI-generated content.
That is why the issue has drawn attention beyond the technical details. For users, the question is not just whether Claude’s answers will change, but what this means for privacy, editing, and the ability to present text as human-written when needed.
The announcement has already prompted strong reactions online. On Reddit, one user described the change as a conspiracy against innocent Claude users, while another argued that people would object only if they wanted to deceive others. Business Insider also reported that “dozens” of users on X said they had canceled their Claude subscriptions because of the move.
Anthropic did not address those reactions in the post. Still, the response shows that many users do not see watermarking as a purely technical update. For some, it goes to trust, control, and whether AI tools should leave a trace in the content they help produce.
Light edits may weaken the watermark without erasing it
One of the main questions Anthropic tried to answer was whether the watermark could simply be edited away. The company said that can happen in some cases, but not all.
Light editing, Anthropic said, will probably not remove the watermark completely. A full rewrite, in which every word is replaced, would remove it.
The company added that if a text has been completely rewritten, it becomes less clear whether it should still be described as AI-generated at all. That suggests Anthropic sees the watermark as tied to the text Claude actually produces, rather than to a later version that a person substantially changes.
The answer gets more complicated when Claude is used only to proofread or lightly edit human-written text. In those cases, Anthropic said detectability will depend on the length of the text and how heavily Claude has edited it. If the changes are minor, nearly all the words would still come from the human author, leaving little for the watermark to attach to.
That sets a practical limit on the system. It is meant to identify text generated by Claude, not to treat every piece of writing that passes through the chatbot in the same way. The more human control there is, the less room there is for the watermark to exist.
Anthropic says code will carry less of a watermark
Anthropic also addressed a concern that matters especially to developers: whether watermarking will affect code.
The company said code should carry less of a watermark than normal text because Claude has less freedom when writing code. Unlike prose, code often has to work in a specific way, which leaves the model with fewer equally valid wording choices.
Anthropic still said watermarking may appear where there is some flexibility, such as comments within code or terms that can be expressed in more than one way. Even then, the company said the effect on the actual code itself would be negligible.
That is a key point for users who are sensitive to any feature that might alter output. Anthropic’s explanation is meant to reassure them that the watermark is aimed more at surrounding text than at the functionality of the code.
Anthropic says other AI developers will follow
Anthropic also said Claude will not be the only chatbot affected. The company said other major model developers have signed the same Code of Practice and will be implementing their own watermarks.
That suggests this is not just an Anthropic decision, but part of a broader shift in how AI-generated content will be handled under European transparency rules. For users, it could mean more AI systems leaving detectable traces in the text they generate, even if those traces remain invisible in the final output.
For businesses and publishers, the development adds another layer to the growing question of how to verify what is human-written and what is machine-generated. Anthropic’s approach is meant to make that easier in principle while keeping the watermark hidden from ordinary readers.
At the same time, the company’s explanation makes clear that the system is not absolute. Light edits may not remove the watermark, but heavy rewriting can. Code is treated differently from prose. And the watermark applies only where Claude has enough room to make choices in the first place.
