Anthropic wants AI-written text to carry a trace of where it came from. Developers, worried that trace could become a scarlet letter for anyone who uses Claude, are already trying to erase it.

The dispute began with Claude models released on or after August 2, which embed an “imperceptible watermark” in their output. Anthropic says the statistical pattern survives copying and pasting, and is intended to support transparency and its commitments under the EU AI Act. Supporters see a useful signal for readers, employers and universities trying to distinguish human work from chatbot output; one developer called it a tool that could help students who cheat “all get caught.”¹

But the same feature has sharpened fears about what a label can imply. Anthropic cautions that a detection result means Claude was likely involved at some point—not that it wrote everything. Freelancers fear client questions over AI-assisted code, while academics and professionals worry that translation, summarising or proofreading could taint work that was principally their own. Critics also argue a watermark could complicate already murky debates over copyright and authorship. Anthropic counters that the mark concerns whether Claude processed content, not ownership or a user’s rights.¹

Within days of the announcement, the resistance became practical. Paris entrepreneur Guillaume Meyer released an open-source tool that strips hidden characters and metadata before rewriting text to disrupt watermark-carrying word patterns. “I am all for content attribution,” he said. “I am against the watermarking technique.”²

Other developers followed with browser-based and Claude-focused removers, as US searches for “AI watermark remover” rose 60% week on week. Their effectiveness remains unverified, but researchers say the basic weakness is clear: text can always be reworded. The EU requires providers to make labels resilient; it does not expressly forbid third parties from trying to remove them. That leaves the central fight unresolved: transparency advocates want a durable signal, while opponents see a blunt mechanism that risks confusing assistance with authorship.²