Anthropic's watermark marks what Claude touched, which is not the same as what Claude wrote

Anthropic started watermarking Claude's output on 2 August, worldwide rather than only in the EU, and has published neither the algorithm nor a detector ^[1][2]. The company's own support documentation says a detected mark means content "may have been processed by Claude" - not that Claude wrote it ^[4]. That single sentence undoes most of what people want a watermark for. The mechanism explains why. When the model picks each word, the watermark uses a secret key to split the reasonable options into two groups and nudges the model toward one of them; a single word proves nothing, but across 500 or 1,000 words a detector holding the key can see the pattern ^[3]. It sees output. It does not see intent, effort, or who did the thinking. So ask Claude to proofread, translate or summarise something you wrote yourself, and the output carries the mark anyway ^[5]. Run the reverse and the signal fails again: short passages, heavy paraphrasing, older models and stripped file metadata all yield clean text that a machine still helped produce ^[6]. Anthropic states both failure directions upfront ^[7]. An operator holding a positive result cannot distinguish a comma fix from a ghostwritten draft, and an operator holding a negative result has learned nothing at all. Regulation makes the mismatch sharper. The EU AI Act does not require marking when a system performs an assistive function for standard editing, and the Commission's own worked example of that is grammar correction ^[8]. Anthropic marks it regardless, an approach Ars Technica described as "nuke it from orbit" ^[9]. The reason is structural rather than ideological: a watermark applied at the model level cannot tell a full draft from a comma, because the output is all it sees ^[10]. The disclosure half of Article 50 is narrower than most people assume. An AI-written novel needs no label, AI-generated marketing copy needs no label, and text informing the public on matters of public interest does need one unless a named, accountable human editor has reviewed it ^[11]. Set that against the marking behaviour and the coverage runs backwards: the grammar fix the Act deliberately exempted gets marked, while a fully synthetic article can reach readers unlabelled because an editor looked at it ^[12]. Non-compliance penalties run to 15m euro or 3 percent of worldwide annual turnover, and the Commission has given itself powers to inspect and fine models directly ^[13][14]. The objections landed within a day and they split cleanly. Investor Bill Gurley argued that if only Anthropic can read the mark, the company becomes "judge, jury, and prosecutor"; Anthropic told Business Insider it will ship a free detection API ^[15][16]. Former Microsoft executive Steven Sinofsky framed the issue as "data retention and your right to private thoughts free of a digital trail" ^[17]. Simon Smith, who runs generative AI at the health agency Klick, asked whether a grammar check would now be flagged as AI-authored - Anthropic's answer is that the mark shows processing, not authorship ^[18][4]. The software trainer John Crickett asked whether marked AI-generated code complicates a copyright claim that requires showing human input ^[19]. On the other side, developer Donn Felker noted that marking helps models avoid training on their own output, the "snake eating itself" problem ^[20], and Aadit Sheth of The Narrative Company argued readers should be able to tell whether words reflect a person's thinking ^[21]. Anthropic frames the change as compliance, saying it is marking output "to comply with the EU AI Act, and other labs are taking similar steps", and that the watermark does not change the meaning, quality or readability of responses ^[22]. Watch the detection API: whether it ships, and whether it reports false-positive and false-negative rates rather than a verdict.