Anthropic’s answer to the AI transparency push is a watermark readers cannot see—but one that could still follow their work into classrooms, client contracts and editorial disputes. The fight is not over whether Claude wrote better prose; it is over what a hidden signal will be taken to mean.

The company says the move follows EU AI Act requirements that took effect on August 2, and it is applying the system worldwide because it cannot yet reliably limit it by region. New Claude models will carry the mark from launch, with older models to follow. ¹

Anthropic’s system uses a version of Google DeepMind’s SynthID-Text approach: rather than inserting hidden characters, it subtly patterns otherwise plausible word choices. The company says readers will see no difference in quality, creativity or readability, and that the mark carries no information identifying a particular user, organization or chat. “Using our key, one can only answer the question ‘What is the likelihood this was partly written by Claude?’” ¹

That limitation is the center of the backlash. Anthropic says a light proofreading pass may produce too few altered words for a mark to register, while more extensive translation, summarization or rewriting can leave a detectable signal. Exact code is less susceptible because there are fewer interchangeable choices; comments may still carry a pattern. ¹

Critics argue the practical distinction will be lost once an employer, professor or client sees an AI detection result. Ars Technica noted that the EU exempts standard editing that does not substantially alter meaning, but warned that a model-level system may still mark material after a comma fix. ² Ben Thompson put the objection more sharply, likening mandatory attribution for a human-directed tool to demanding that “a ballpoint pen advertise itself as the author.” ³

Some Claude subscribers have already canceled. Users told Business Insider they worried that invisible signals in code, translations or client documentation could invite compliance penalties or authorship challenges even when their underlying work was their own. “Even if you edit your own work that you wrote, they add the watermark anyway,” one consultant said. ⁴

Anthropic counters that a detected mark means Claude processed content, not that it originated it, and says it will offer a free detection API. Yet critics question whether that nuance can survive real-world use—and why a tool meant to expose synthetic content may attach suspicion to ordinary human work. ⁵ Supporters see the trade-off differently: watermarking could help audiences distinguish AI writing and curb the recycling of AI output into future model training. ⁶