EU's text watermarks won't stop disinformation, experts warn
Get the Health newsletter
Daily health & science — research, biotech, public health, the studies worth knowing. Free.
- EU AI Act watermarking requirement for AI-generated text took effect 2 August, with all deployed models required to comply by 2 December
- OpenAI already watermarks images and audio and plans to add text watermarking, while Anthropic has announced future Claude models will include text watermarks
- James Padolsey at AI safety firm NOPE calls text watermarks "not useful," noting they can be removed by light edits and bypassed via open-source models outside any firm's control — he built the declaude tool to prove it
- Peter Scarfe at the University of Reading warns of false positives, noting students who merely asked AI to grammar-check their essays could be wrongly flagged as having AI-generated text
- Erman Ayday at Case Western Reserve University argues even trivially removable watermarks add friction, since defeating them takes effort that may equal simply writing the content from scratch
- A European Commission spokesperson defended the rules, stating the adversarial robustness of marking and detection solutions must be assessed for resilience to copying, removal, and modification attacks
Why it matters: For the EU's AI Act watermark mandate to function, tools must resist tampering while avoiding false flags against legitimate AI-assisted work. With experts saying text watermarks are removable by light edits, bypassable through open-source models, and prone to misidentifying students who merely grammar-checked essays, the 2 December deadline risks delivering transparency in name only while educators absorb the false-positive costs.
Ask SkimNews
