Paying users are abandoning Anthropic Claude due to the appearance of “watermarks” in the generated text
17 August 19:11
Last week, Anthropic announced that it is implementing a system to label and identify content generated by the Claude chatbot. Embedded watermarks and digital metadata will be added to texts and files across all of the developer’s platforms, according to [Komersant].
These changes are being implemented as part of compliance with the EU Code of Practice on AI Transparency. Claude models launched in the European Union starting in August 2026 will support machine-readable labeling by default, while the feature will be added to older versions at a later date.
The watermark is woven directly into the text’s structure. It remains invisible, does not affect the quality or content of responses, is preserved when copied, and can withstand partial editing. The company is also developing tools that will allow third parties to verify the presence of such marks.
Watch us on YouTube: important topics – without censorship
At the same time, Anthropic noted that the presence of a watermark does not always indicate that the text was entirely generated by artificial intelligence, as users often use Claude only to edit, translate, or format their own ideas.
In addition, watermarks may disappear in cases of extensive rephrasing, significant text shortening, or the removal of metadata when changing the file format.
Following this announcement, paid Claude users began canceling their subscriptions after Anthropic announced that it would no longer mark texts generated by the chatbot. Some people stated on the social network X that they were giving up paid access out of concern that others would be able to detect the use of artificial intelligence, according to a report by Gizmodo.
One user, who canceled his subscription to the Claude Max plan—which costs $100 per month—had been using the AI assistant to write applications and other tools. He is concerned that a watermark might appear even if the text was original and Claude was used only for translation, correcting academic citations, or creating a resume. The watermark poses problems in academic and professional settings, where strict rules regarding AI are in place. After canceling his subscription to Anthropic, this customer chose Cursor and Grok 4.6 for writing code and Grok Bot for work and personal tasks.
Another user who canceled their paid subscription said they use AI for code review and translation, and fears that an AI watermark would be present in the code sent to clients, raising concerns about authorship and potential contractual penalties. This user opted for Chinese AI providers. A third user said he uses Claude to create business documentation—he is concerned about the presence of AI watermarks in texts for which he wrote the source material himself, while the chatbot merely edited it.
A fourth user, who canceled his subscription, works as a software engineer—he canceled his Claude Max subscription back in July. In his opinion, the quality of service on the platform has declined, access to the best models has become limited, and more reliable alternatives have emerged. The watermark incident only confirmed that his decision was the right one: he did not want the code he had written as a professional to raise questions about authorship, compliance with requirements, or client policies.
In addition to Anthropic, AI content is labeled by Google, which uses its own SynthID technology, and the AI assistant on Elon Musk’s social network, X. The latter’s watermark recently appeared briefly in one of the White House’s official statements. Anthropic stated that it added the watermark to comply with new EU regulatory requirements and promised to release a free API to check content for the presence of the Claude watermark. One of the company’s now-former clients stated that its new initiative only reinforced his concerns about excessive dependence on Claude. This is typical of proprietary software, he acknowledged—a developer can change anything at any time. Therefore, the AI used in one’s work should not depend on a single provider.
Read us on Telegram: important topics – without censorship
