Anthropic to Embed Invisible Watermarks in Claude Text Globally Under EU AI Act
TinfoilTricorn · x · 2026-08-12
To comply with the EU AI Act, major Western AI labs including OpenAI, Google, Meta, and Anthropic have signed a code requiring AI-generated text to be machine-readable marked.
Anthropic specifically announced it will embed invisible watermarks in all generated text for new Claude models globally. The watermark is part of the text itself rather than metadata, persisting through copying, pasting, and some editing. Models launched after August 2, 2026, must include this, with existing models needing compliance by December 2, 2026.
Notably, Anthropic is applying this EU requirement worldwide, even though no US federal law mandates it. Google and OpenAI are also expanding provenance signals across modalities.
More from Models
- LlamaIndex Launches ExtractBench: 4,869 Pages of Complex Docs Across 8 Domains — llama_index · 2026-08-12
- Safety Guardrails Hinder Bug Fixing Due to Keyword Triggers — xeophon · 2026-08-12
- Mitsuhiko Asks: Will Closed SOTA Labs Ban Assistant Prefill? — mitsuhiko · 2026-08-12
- Encrypted CoT Vulnerability Leaks API Keys; NVIDIA Launches Nemotron 3.5 Lightning — Latent Space · 2026-08-12
- Anticipating Grok 4.6, GLM 5.5, and DeepSeek V4 Pro Releases — bindureddy · 2026-08-12
- Meta's Muse Glimmer 30B Updates Chat Template — TacticalRock · 2026-08-12