Anthropic to Globally Deploy Invisible Text Watermarks for Claude to Comply with EU AI Act
新智元 · wechat · 2026-08-11
In response to the transparency requirements of the EU AI Act (Article 50), Anthropic has announced it will tag outputs across all Claude products globally. This marks the first large-scale deployment of text watermarks by a mainstream AI company.
Two-Layer Tagging Technology
The first layer is a Text Watermark: created by introducing an invisible statistical bias during Token generation. This watermark persists through copy-pasting, and the company claims it does not affect text quality or readability. The second layer is File Provenance Metadata: attaching C2PA digital signatures to generated files (like images) to mark origin and detect tampering.
Limitations and Controversies
Anthropic cautions that the watermark is not foolproof: its presence doesn't mean Claude is the sole original author (e.g., if used for translation or polishing), and heavy rewriting or very short texts can render it undetectable. Conversely, the absence of a watermark doesn't guarantee human authorship. This move has sparked concerns among users regarding source neutrality, with some paid subscribers needing untraceable content expressing dissatisfaction. This could indirectly boost the appeal of open-source models.
More from Models
- First Model to Pass the Video Turing Test: Half of Users Thought It Was Human — Puzzleheaded-King584 · 2026-10-03
- Using System One models in Swift: fast, deterministic decisions via Apple Foundation Models — rxwei · 2026-10-03
- Linux Kernel CVEs Surge From ~500 to 1500+ Per Release, LLMs Blamed for Bulk of the Rise — burny_tech · 2026-10-03
- Developer Complains OpenAI's Coding Model Endlessly Scopes Creeps Instead of Finishing Tasks — DavidWells · 2026-10-03
- Sonnet 5 Spotted in Google Antigravity Backend, Which Still Runs Sonnet 4.6 — brandon_galang · 2026-10-03
- Post-training Yandex AliceAI-80B-A3B from scratch: a NaN bug in custom V100 kernels killed one run — jjusko20 · 2026-10-03