AI Text Watermarks Defeated by Light Edits, Experts Warn
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- EU AI Act watermarking requirements for AI-generated text, images, and video took effect August 2, with all models required to comply by December 2
- OpenAI already watermarks images and audio and plans to add text watermarking; Anthropic announced future models will watermark text
- James Padolsey at AI safety firm NOPE says text watermarks can be deleted by light edits, built an online tool called "declaude" to strip them, and warns open-source AI models outside EU control will remain available to bad actors as cheaper alternatives
- Peter Scarfe at the University of Reading warns of false positives — students whose work was only lightly processed by AI for grammar checks could be wrongly flagged as AI-generated
- Erman Ayday at Case Western Reserve University argues even trivial watermark removal creates friction that could deter mass disinformation campaigns over time
- The European Commission acknowledged adversarial robustness must be assessed for resilience to copying, removal, regeneration, and modification attacks on the markings
Why it matters: University students face false-positive risk — work could be flagged as AI-generated when AI only helped with grammar checks. Meanwhile, Padolsey's "declaude" tool strips watermarks in seconds, meaning the EU mandate burdens legitimate users while failing to deter bad actors who switch to open-source models outside EU jurisdiction.
Ask SkimNews
