An article generated by AI doesn't require a label. However, your reviewed email does.

      Begin with what the mark truly is, as it is not merely an imprint on the page. When a model generates text, it selects each word from a variety of plausible choices. A watermark discreetly separates these options into two categories using a confidential key, gently guiding the model towards one of the categories. A single word does not confirm anything. However, across 500 or 1,000 words, a detector with the key can detect the pattern.

      Anthropic started marking Claude's output on August 2, rolling it out globally instead of just in Europe. The company has yet to release the algorithm or a detection tool. It indicates what Claude interacted with, rather than what Claude authored. This crucial detail has often been overlooked in coverage, with Anthropic making it clear in its support article. A detected mark signifies that the content “may have been processed by Claude”, but does not imply that Claude authored it. If you ask the model to proofread, translate, or summarize your own text, the output will still carry the mark.

      Conversely, the absence of a mark does not indicate no AI involvement. Short texts, extensive paraphrasing, older models, and files with stripped metadata can all yield clean text that was still influenced by a machine. Thus, the signal has two potential pitfalls, which the company acknowledges upfront.

      The grammar fix is exempt from the law. Here is where things become peculiar. The EU AI Act does not mandate marking when a system aids in standard editing functions, and grammar correction is cited as an example by the Commission. Despite this, Anthropic applies a mark anyway, leading Ars Technica to describe the strategy as “nuke it from orbit.” This decision stems from structural reasoning rather than ideological. A watermark applied at the model level cannot differentiate between a complete draft and an individual punctuation mark; it marks the output, which is all it can perceive. Consequently, the mark appears on content that the law was intended to leave unmarked.

      Now, let’s examine the other aspect of Article 50, which regulates what publishers are obligated to inform readers. This requirement is much narrower than most believe. An AI-generated novel is not required to be labeled. AI-generated marketing text does not need a label either. However, text that informs the public on matters of public interest does need a label unless a named and accountable human editor has reviewed it. When combining these two elements, a contradiction emerges: every keystroke receives a watermark at the model level, while a completely synthetic article can be presented to readers without a label as long as an editor reviewed it.

      The EU has been developing this framework for some time. It has already established mandatory labels for synthetic content and has granted itself powers to inspect and impose fines on models directly. The penalty can reach €15 million or 3% of annual global turnover.

      Objections surfaced within a day. Investor Bill Gurley argued that if only Anthropic can interpret the mark, it essentially becomes “judge, jury, and prosecutor.” Anthropic responded to Business Insider by stating that it will provide a free detection API for anyone to verify. Former Microsoft executive Steven Sinofsky raised another concern, emphasizing issues around “data retention and your right to private thoughts free of a digital trail.” Simon Smith, who oversees generative AI at the health agency Klick, inquired whether grammar checks would now be identified as AI-authored. Anthropic’s response remains consistent: the mark indicates processing, not authorship.

      Trainer John Crickett posed a pertinent question regarding code: if AI-generated code carries a mark, does this complicate copyright claims where the author must demonstrate human input?

      Nonetheless, there is a case for marking outputs. Developer Donn Felker pointed out that marking helps prevent models from training on their own output—a situation he referred to as a “snake eating itself.” Aadit Sheth of The Narrative Company emphasized the importance for readers to recognize whether the text reflects an individual’s thoughts. The Commission’s rationale is even broader, aiming for individuals to adjust their trust based on perceived risks of fraud, impersonation, and consumer deception.

      Anthropic presents this as a matter of compliance. The company stated it is adding marking “to comply with the EU AI Act, with other labs taking similar actions.” It further claims that the watermark does not alter the meaning, quality, or readability of Claude’s outputs.

      Several questions remain unanswered. Ars Technica requested a timeline for detection, inquired about testing on false positives and negatives, and how the marks relate to the standard-editing exemption. Anthropic’s response did not address these inquiries. The false-positive rate is particularly significant, as it determines the weight a teacher or editor should attribute to the mark.

      The uncomfortable reality is that the mark may flag legitimate content while overlooking deliberate manipulation. If watermarked text is pasted into a second model for a rewrite, the signal is likely eliminated. File-based provenance is even less reliable; a screenshot will erase it, as will