The EU just forced AI companies to watermark their output, and the watermark remover already exists as a one-click web app.
The Summary
- Anthropic plans to watermark all Claude-generated text to comply with new EU regulation, but the marks are trivially removable through simple rewriting
- Developer James Padolsey argues the law treats AI assistance differently than calculators or spellcheckers, punishing capability rather than misuse
- Declaude, a free web tool, already strips AI watermarks with one click, exposing the fundamental weakness of compliance-first watermarking
The Signal
Anthropic's implementation goes beyond the EU law's minimum requirements. The regulation allows exemptions for assistive use cases. Anthropic chose a blanket approach anyway. Every Claude output gets marked, whether you're drafting a phishing email or helping someone with dyslexia compose a cover letter.
Padolsey's interactive essay breaks down exactly how these watermarks work: they nudge the model toward specific word choices during generation. The signal is subtle enough that humans rarely notice. It's also subtle enough that motivated actors remove it by asking any other LLM to rephrase the text.
"The result is a signal broad enough to implicate harmless and assistive use, yet fragile enough to be removed by a motivated person through substantial recomposition."
The irony runs deeper. Declaude was originally built to remove Claude's personality quirks, the "delighted to help" and "I appreciate your question" verbal tics that flag AI-generated text more reliably than any technical watermark. Now it serves double duty as a watermark stripper. One paste, one click, clean output.
This creates a detection asymmetry that makes the regulation worse than useless:
- Assistive users get flagged for using accessibility tools
- Non-technical users get caught by automated detection
- Bad actors remove watermarks trivially and continue
Google Gemini already implemented similar watermarking. The pattern is clear: tech companies will comply with regulations that sound good in Brussels but break in contact with reality. The alternative, fighting the regulation or refusing EU market access, carries higher cost than shipping theater.
The calculator comparison hits hardest. Nobody demands that calculator output be marked as "machine-assisted." Spell-check doesn't append metadata tags. The boundary the EU drew is arbitrary. It penalizes capability, not intent. A tool smart enough to write a full sentence gets treated differently than one that corrects grammar, even when both serve identical functions for users who need help writing.
The Implication
Watch what happens when detection fails and suspicion spreads anyway. Students, accessibility tool users, and non-native speakers face the brunt. Meanwhile, the people the regulation aimed to catch just run their output through a paraphraser.
The real signal here isn't about watermarking. It's about regulatory capture by compliance complexity. Laws written by people who don't understand the technology they're regulating create work for lawyers and engineers while solving nothing. The EU wanted to stop AI-generated deception. They built a system that flags help-seekers and misses deceivers. That's not a bug in implementation. That's the law working exactly as designed by people who confused difficulty with integrity.
If you're building tools in this space, the takeaway is clear: assume watermarks are performance art for regulators, not actual security. Build for the world where users need help and bad actors remove detection signals at will. That's the world we already live in.