Anthropic adds text watermarking to Claude, raising quality and legal concerns
The Decoder · rss · 2026-08-17
Anthropic has implemented text watermarking for its Claude models to help identify AI-generated content.
Technical Mechanisms and Criticism
- The watermark works by subtly adjusting word choices to embed a detectable signal.
- Critics argue this could degrade text quality or fluency, and its effectiveness across non-English languages is uncertain.
Regulatory and Legal Implications
- The move aligns with EU AI Act requirements for labeling AI content.
- Legal experts worry that watermarking might impose constraints on expression, creating new transparency challenges.
More from Safety
- How to gate Agent actions in production environments? — Excellent-Park-1160 · 2026-08-17
- New Dataset Aligns NIST RMF with AI Governance Standards — iamKierraD · 2026-08-17
- AI Policy Debate Needs Clear Categorization of Use Cases — emollick · 2026-08-17
- Anthropic says its AI models hacked 3 orgs during testing — Traditional_Blood799 · 2026-08-17
- EU firms may use Chinese open models via "jurisdictional wrapper" — teortaxesTex · 2026-08-17
- ChatGPT has quietly built a profile on you — 15 prompts to see and wipe it — LearnWithBishal · 2026-08-17