Anthropic Watermark Terms Exposed: User Text Marked, Liability Capped
arnoldwender · reddit · 2026-08-12
A German commercial user who read Anthropic's actual contracts highlighted four critical issues missing from recent watermark coverage:
- Marks can hit user-provided text: Anthropic's support docs admit that even if users just ask Claude to proofread their own writing, the output can carry the AI mark. This over-applies the EU AI Act, which targets AI-generated content.
- Terms never mention marking: Both Commercial and Consumer Terms predate the announcement and lack any clauses authorizing watermarking or C2PA metadata, despite assigning output ownership to the user.
- Legal action is economically dead: EEA users are bound to Irish law and arbitration in Dublin. Class actions are waived, and liability is capped at the previous 12 months' fees—making individual arbitration financially unviable.
- Chinese models offer no escape: China's equivalent AI labeling measures have been in force since September 2025, applying to models like DeepSeek and Kimi.
The author advises affected users to file regulatory complaints with national AI market-surveillance authorities before canceling, specifically targeting the misapplication of marks to user-supplied content, which bypasses arbitration clauses.
More from Safety
- Satirizing AI Copyright Extremism: 'Stealing' Rhetoric Spreads to Hacker News — voooooogel · 2026-08-12
- ChatGPT Reasoning Trace Extraction Vulnerability Patched — wunderwuzzi23 · 2026-08-12
- GuideLabs Releases Interpretable LLM with Outputs Traceable to Training Data — andreas_madsen · 2026-08-12
- AI Labs Got Complacent on Alignment as RL Scaling Raises Rogue Risks — Afinetheorem · 2026-08-12
- Anthropic Sparks Backlash by Watermarking All Claude Code Outputs — cjimti · 2026-08-12
- Long Benign Context Passively Decouples RLHF Alignment Without Jailbreaks — PresentSituation8736 · 2026-08-12