Anthropic has begun embedding an invisible watermark in texts generated by its Claude model. No hidden characters are added to the content, yet with a sufficiently long passage, it is possible to mathematically verify the probability that it was authored by AI. The company took this step because of European legislation. However, some users see it as a betrayal, and some are even canceling their subscriptions.
How word choice becomes a signature
To understand the entire principle, it is crucial to know how large language models work. Claude generates text word by word, choosing from several of the most suitable candidates for each subsequent position. In the sentence “Today’s weather was cold and…” the system is highly unlikely to choose the word “sweet,” opting instead for something like “overcast” or “gray.” Readers generally do not care about the specific choice, and the final result is determined by an algorithm that generates random numbers.
And this is exactly where Anthropic intervened. The source of this randomness is no longer a conventional generator, but a specific key combined with several preceding words. The choice remains random, but anyone who possesses this key can retrospectively verify whether the sequence of words corresponds to what the model would have selected based on the key. This creates a pattern in the text that readers have no chance of noticing.
Anthropic explained this using the example of Monopoly. Instead of rolling dice, players would take values from a book full of the digits of pi, starting at a random point and then proceeding sequentially. Nothing changes in terms of how the game itself unfolds, but anyone familiar with pi can determine after the game that the moves were not decided by conventional dice.
Where the mark appears and where it does not
However, the method has its limits. The key can only be used to determine the probability that Claude contributed to the text. It cannot confirm that a text was written by a human, nor can it detect another artificial intelligence, since that AI would use its own key or a completely different method. Detection also works poorly with short passages because they contain few decisions, while confidence increases as the text becomes longer.
The method also fails where there is only one correct option. After the words “Newton’s most famous work was called Principia…” the next word simply has to be “Mathematica,” so the watermark has nothing to work with. The same applies to calculations, where there is only one correct answer after the prompt “2 + 2 =”.
Programmers will be interested in how this applies to source code. It carries the mark more weakly because the model must write a functional solution and often does not have many alternatives to choose from. The watermark may appear in places where the choice is arbitrary, such as comments, but its presence in the functional program itself is negligible.
Editing someone else’s text is a special case. The watermark applies only to expressions selected by Claude itself. When the model merely makes light corrections to a text, most of the original words remain human-written, and the mark has little to attach itself to. With translations, however, the exact opposite is true. There, the model chooses every word, so the resulting text contains the watermark.
The reason for introducing watermarks
The company admits that it is fulfilling an obligation. In July 2026, together with other major AI model providers and approximately 190 entities, it signed the European code of practice on transparency. This requires system developers to clearly label machine-generated content. Under EU rules, all texts created by artificial intelligence must be labeled in this way from December onward.
Anthropic watermarks texts globally because it does not yet have a reliable way to restrict the feature to a specific region. Claude will not be alone in this, either. From August 2, the EU requires all providers operating in its market to label content. Other technology developers have also signed the same commitment and are already preparing their own watermarks. Although older models launched before this date are subject to a transition period, Anthropic plans to retrofit them in the coming months.
Will text quality deteriorate?
The most vocal expert objection concerns output quality. Well-known blogger John Gruber described the entire approach as destroying the very essence of writing. He claims that the watermark ties Claude’s hands and forces it to choose less suitable words. According to him, the result may not become outright incorrect, but it will still be worse overall.
Steven Murdoch, a professor of computer science at University College London, sees it differently. According to him, the change will most likely not be noticeable at all. Randomness is already essential for these models, as otherwise they would get stuck in loops and constantly repeat the same thing. The random number generators remain the same, but their output is now statistically predictable. He also pointed out that the model does not actually think about choosing between two synonyms; it simply generates them based on randomness.
Anthropic relies on its own measurements. In its own testing, it found no negative impact on content, creativity, or readability. It refers to research by Google DeepMind, where the developers of the SynthID-Text method deployed a watermarked model in Gemini’s production environment. They found no statistically significant difference in user ratings based on thumbs-up or thumbs-down feedback. Furthermore, in a comparative study, human evaluators detected no qualitative difference between watermarked and unwatermarked responses.
What users are afraid of
Reactions on social media were fierce. One Reddit post called the move a conspiracy against innocent users, while another countered that the only reason to reject the watermark was an attempt to deceive people. Business Insider also reported that dozens of users on X had already announced that they were canceling their subscriptions because of the introduction of the mark.
According to critics, more advanced users will get around it by rewriting the text or having another service modify it, while ordinary people will easily give themselves away. Examples include students who have a paragraph rephrased, journalists requesting a summary of a two-hundred-page transcript, or authors with writer’s block looking for synonyms. TechCrunch, however, argues that these concerns are unfounded. A generated summary poses no threat to a journalist unless they copy it directly into their article, which they should not be doing anyway. According to the publication, the same applies to students and their essays.
The second line of resistance revolves around authorship. One dissatisfied user called this hidden labeling unethical. They argued that they had provided the instructions, context, decisions, and countless revisions themselves, meaning Claude was merely a tool. They also asked why the model was claiming any credit at all. Others countered that credit was not the issue. The goal is merely to make it possible to identify machine-generated content, because it can cause harm in certain situations. Another critic chose a more sophisticated argument about hypocrisy. They pointed out that the models themselves were created by ingesting vast amounts of other people’s copyrighted work.
Anthropic tempers reactions
Anthropic itself is tempering expectations about what the watermark can prove. It only shows that Claude was probably involved in creating the content in some way. However, it cannot distinguish whether the model wrote the text entirely or merely rewrote it substantially. Moreover, there is no trail leading to the user or their organization, and neither the mark nor the key contains any data from which their identity could be traced. This does not change ownership of or responsibility for the resulting output in any way.
The practical side of the matter remains unresolved. The company is still developing an interface for verifying the watermark and must also finalize the details of its ultimate form. The situation is entirely different for multimedia files. Images and other supported formats receive a cryptographically signed mark in their metadata according to the open C2PA standard. This is used by both camera manufacturers and photo editors, while nothing in the file’s actual content changes.
It is possible to remove the mark. Minor edits will probably not erase it completely, but a total rewrite that leaves no stone unturned will. In that case, however, the question arises as to whether the resulting text can still be considered the work of artificial intelligence at all.
Sources: theguardian.com and techcrunch.com
