Anthropic's Invisible Watermarking Criticized as Brussels Rule Goes Global
r0ck3t23 · x · 2026-08-17
Anthropic has begun embedding invisible watermarks in Claude's output, a compliance measure for the EU AI Act that is now applied globally by default with no opt-out option.
Mechanism and Impact:
- Watermarks are embedded in word choices; heavier reliance on the model (e.g., non-native speakers translating) leaves a stronger signal.
- Local models (like Llama) are unaffected, creating an inequity where paying API users get flagged while open-source users do not.
- A planned detection API will allow schools and employers to classify text instantly, risking false positives and lack of recourse.
- Tools designed to strip these marks have gained thousands of stars overnight, signaling strong user resistance.
The piece criticizes this unauthorized global regulatory overreach, arguing it measures "obedience" rather than ensuring safety.
Related event: Anthropic Adds Invisible Watermarks to Claude, Sparking Global Backlash(21 posts)→
More from Companies & People
- Radical Numerics opens 19 roles including Head of BD to build next-gen biological AI — exnx · 2026-10-03
- Cohere VP of Engineering Rachad Alao to keynote PyTorch Conference in San Jose, Oct 20-21 — PyTorch · 2026-10-03
- Ray Dalio on "dotting": real-time peer feedback in meetings — RayDalio · 2026-10-02
- Gergely Orosz: startups adopted 'hip' tech partly to attract blocked senior devs — mgill25 · 2026-10-02
- G2i hiring 50 staff engineers at $200-300/hr to build evals for frontier AI labs — threepointone · 2026-10-02
- Every's Thesis: 2027 conference on AI and work lands in Brooklyn on Nov 5 — every · 2026-10-02