Anthropic Adds Watermark to Claude, Cancellations Follow

Anthropic introduces an "imperceptible" watermark in texts processed by Claude, applied globally since August 2. Several professional users report having canceled their subscriptions, fearing an AI marking on delivered code or documents. The company cites compliance with the AI Act and announces a free verification API, while claiming not to have seen an increase in cancellations.

Anthropic targets the AI Act with announced verification API

Anthropic states that it has integrated the watermark to comply with European AI regulations and announces the upcoming availability of a free API that allows both users and third parties to verify for themselves whether a text carries this marking. In the ecosystem, other companies are already flagging AI-generated content: Google uses SynthID, and OpenAI employs it for images as well as supported audio. The platform X labels content it deems generated or manipulated by AI with a "Made with AI" tag. This mention appeared and then disappeared from a resignation post by Karoline Leavitt on Wednesday. Proponents of watermarks argue for transparency, while others see it as a way to limit the training of models on content already generated by AI. On Tuesday, Aadit Sheth publicly defended the idea that the public should be able to identify whether words reflect a person's thoughts, citing watermarks as a means to achieve this.

Anthropic describes an imperceptible watermark that persists

Anthropic has indicated that certain versions of Claude would incorporate a so-called "imperceptible" watermark in their text outputs. According to the company, this marking accompanies the content during a copy-paste, can survive certain modifications, and aims to distinguish AI-generated text from human writing. The company clarifies that the watermark does not alter the meaning, quality, or readability of the responses. It also states that a detected watermark means that Claude has processed the text, without necessarily being the author, and that the mark can remain after proofreading, translation, or summarization. Anthropic has set the global application of this mechanism for supported models to August 2. Dario Amodei publicly announced the addition of this marking and its possible persistence despite edits.

Professionals fear marking of their code and documents

Several professional use cases are concerned about the prospect of a text being marked as having passed through AI. One developer expressed concern that an AI label on delivered code could raise doubts about authorship. Some users fear that the use of AI will be revealed in deliverables submitted to clients, particularly in the software development field. Vladislav Rajtmajer fears that a marking could lead to questions about authorship or even trigger contractual penalties. Richard Echols, who writes documentation for companies, expresses his concern about the marker appearing even when he has written the content himself and only consulted the chatbot for corrections. Stevan Bogosavljević believes that code attributed to Claude would be undesirable and refuses to accept that an invisible marker could raise questions about compliance or client policies. "Even if you modify your own work that you wrote, they add the watermark anyway," Richard Echols stated.

Subscribers cancel, others switch to alternatives

Claude users report having canceled their subscriptions due to the watermark, with dozens posting about it on X since Monday, some providing confirmations. Many say they refuse to allow a persistent marker into their production. On Wednesday, Arturo Villarroya canceled his Claude Max subscription, a service starting at $100 per month. He says he primarily uses Claude for coding applications and tools, but also for research, corrections, and checks. He fears that a watermark could persist even when the writing is original and Claude was only used for translation, spelling corrections, adjusting academic citations, or slight summarization, which could pose problems in environments with strict AI rules. He indicates he is turning to Cursor and Grok 4.6 for coding, and Grok Bot for organization. On Tuesday, freelance developer Vladislav Rajtmajer also canceled, citing the marking as the main reason. He uses AI for code proofreading and translations and sees the watermark as a signal to test Chinese AI providers and other models. Richard Echols, a consultant in Georgia, did the same on Wednesday, indicating that the watermark influenced his decision; he uses it for corporate documentation. In July, senior engineer Stevan Bogosavljević had already canceled, citing declining service quality, limited access to the best models, and stronger alternatives; he believes the watermark reinforced that choice. This week, engineer and founder Santiago Alessandro Rivera Martínez also canceled, arguing that a flow based on proprietary software can see its rules, features, restrictions, or conditions change. He prefers a portable AI layer independent of the provider.

Anthropic reports no increase in cancellations

Anthropic claims not to have observed an upward trend in cancellations since the announcement of the watermark. It is also considered too early to assess the potential impact on the company. The number of professional clients was 300,000 in September 2025, a volume presented as likely having increased since, with no more recent numerical indication available immediately.