Claude's invisible watermarks cracked within hours; override code gets 20k bookmarks
deliprao · x · 2026-08-24
Anthropic announced last week that Claude would globally embed invisible, machine-readable watermarks in AI-generated text to comply with the EU AI Act. Within four hours, developer Guillaume Meyer published his override. The watermark-removal code went viral on GitHub, drew over 20,000 bookmarks on X and 100+ contributors, with many folding the technique into their own projects.
The new rules require providers like Anthropic and OpenAI to label synthetic audio, image, video, or text so machines can detect it as AI-generated, on pain of fines up to 3% of annual turnover. Meyer says some evade the watermark on principle — opposing mandatory labeling of all AI content — while others, himself included, relish the technical challenge; freelance writers and social creators have asked him for help using the code.
Researcher Delip Rao quoted the story to argue the literature covered this long ago, and that WIRED should have written about how out-of-touch EU AI regulators are — with Anthropic researchers knowing this "safety" feature was dead on arrival and shipping it purely as bureaucratic compliance.
More from Models
- Unredacted Anthropic filing exposes internal data-quality debate months before Opus 3 cutoff — zacharynado · 2026-10-15
- Google says teams across the company are using Gemini 4 Argon on its hardest engineering problems — CurieuxExplorer · 2026-10-15
- Perplexity's new embedding models: index with a 9B model, search with 0.6B — antoine_chaffin · 2026-10-15
- 797GB of n-gram counts in CPU RAM let 1B models match 57% more compute — PangWeiKoh · 2026-10-15
- Leak: Moonshot's Kimi flagship 'x4' done with pre-training, K3.1 in internal testing — lyraxana · 2026-10-15
- Unsloth ships GGUF quants for Nemotron 3.5: Bolt-74B runs in 24GB RAM — danielhanchen · 2026-10-15