EU AI Act's Watermarking Mandate Criticized as Impractical and Harmful to Model Performance
Dan_Jeffries1 · x · 2026-08-11
AI researcher Dan Jeffries criticizes the EU AI Act as one of the worst pieces of legislation ever written, arguing that its watermarking requirement is based on magical thinking. He explains that text watermarking cannot be done without altering the text, and methods like changing word choices or inserting hidden Unicode characters are easily defeated by editing. Thus, the mandate forces models to write in statistically worse, more uniform ways, harming performance across tasks, and benefits big tech by imposing higher compliance costs on smaller competitors.
More from Safety
- French News Publishers Sue Google Over AI Overviews Days After Launch — gaganghotra_ · 2026-08-12
- Researcher Explains LLM Watermarks: Uses Minimal Entropy, Fails on Low-Entropy Outputs — RyanGreenblatt · 2026-08-12
- AI Watermarking Limits: Low-Entropy Outputs Missed, But Social Benefits Outweigh Costs — RyanGreenblatt · 2026-08-12
- Vs. Pangram Detector: AI Watermarking Limitations on Human-Sourced Edits — RyanGreenblatt · 2026-08-12
- Researcher Predicts Claude Watermarking Will Fail Due to False Positives — joshgans · 2026-08-12
- Claude Officially Marks AI Content Steganographically, False Positives Reported — johnnyApplePRNG · 2026-08-12