Anthropic’s push to make Claude-generated text traceable was meant to strengthen transparency. Within days, it had produced the opposite kind of signal: a booming market for ways to make those traces disappear.

Since August 2, Anthropic has embedded imperceptible statistical watermarks in text from supported Claude models, saying the measure supports its commitments under the EU AI Act. The label is designed to persist through copying and pasting, but critics say that durability is precisely the problem when AI has been used only to edit, translate or summarize a largely human document.

The backlash quickly became practical. Search interest in “AI watermark remover” rose, while developers released tools intended to rewrite text, strip metadata or otherwise disrupt the patterns detectors rely on. One such project, Guillaume Meyer’s open-source Watermarks Remover, gained more than 14,000 GitHub stars, although its effectiveness has not been independently established. Researchers warn the broader contest is inherent to the technology: “There will always be ways to remove the watermark,” ETH Zurich researcher Thibaud Gloaguen said, including simply rewording the text.¹

Meyer says he built an initial version in roughly five hours after Anthropic’s announcement and watched it go viral after an August 11 post on X. His objection is not to disclosure itself but to a system that may brand work as AI-made after minimal assistance. “I am all for content attribution,” he said. “I am against the watermarking technique, and that’s a very significant distinction.”²

That distinction may soon matter beyond an online skirmish. EU guidance asks AI providers to make markings resilient to alteration and adversarial attacks, yet the AI Act does not expressly prohibit third parties from building removal tools. Legal experts cited by Business Insider say the sharper risk falls on users who erase a label to pass AI-generated work off as human-made—conduct Anthropic’s policies prohibit.

Anthropic, which did not answer questions about the removers, plans a text-detection API alongside a future model. By then, the company may be confronting an old watermarking reality in a new format: every public way to identify a mark can also help someone learn how to defeat it.