2 weeks ago
Anthropic Adds Invisible Watermark to Claude Text Worldwide
Anthropic, the company that makes the Claude chatbot, has started hiding a special secret mark inside the words Claude writes.
The mark is invisible, so people cannot see it when they read the text.
It stays in the text even after copying and pasting it into other documents.
The mark lets computers know the words were written by a machine and not a person.
Computer-made files also get a digital certificate that shows how they were made.
The company is adding the mark everywhere in the world, not just where the law requires it.
The rule comes from the European Union, which says machine-made content must carry a label.
The mark can help schools, news websites and online platforms spot text written by computers.
But the mark cannot catch every computer program, because some free AI programs can remove it.
So someone who really wants to hide machine-made writing can still do it.
Anthropic will embed an imperceptible watermark directly into text generated by its models to identify machine-produced content.
The watermark is invisible to readers and designed to survive copying, pasting, reformatting and moving between applications; for generated files, Anthropic uses signed provenance metadata based on the C2PA standard.
The marking is being applied worldwide, not only in Europe, even though the legal obligation comes from European rules.
Claude models launched in the EU on or after August 2, 2026 carry the marking from launch, and earlier models are being updated to support it.
The watermark aids attribution at scale but cannot stop determined misuse, since open-weight models like Moonshot's Kimi K3, Meta's Muse Glimmer, and DeepSeek's V4 Flash fall outside the scheme.
- Who
- Anthropic, the maker of the Claude AI models, responding to European regulation
- What
- Adding an invisible watermark to Claude-generated text and C2PA provenance metadata to generated files to label AI-produced content
- Where
- Worldwide, including outside the European Union
- When
- From August 2, 2026, when the EU AI Act's transparency obligations became enforceable; older Claude models are being updated to support the marking
- Why
- To comply with Article 50 of the European Union's AI Act, which requires AI-generated content to be labelled as machine-made
Watermarking as a global standard
Watermarking as a limited compliance measure
Scope of application
Watermarking as a global standard
Applying the marking worldwide builds one compliant product and echoes how the General Data Protection Regulation reshaped privacy practice far beyond European borders, showing a rule written in Brussels can change products for users everywhere.
Watermarking as a limited compliance measure
Anthropic could have marked output only for EU users, since the legal obligation is European, and maintaining separate marked and unmarked pipelines would avoid engineering overhead for the rest of the world.
Effectiveness against misuse
Watermarking as a global standard
The watermark gives platforms, publishers and educational institutions a technical means to attribute machine-generated text at scale, a meaningful capability against fabricated content such as deepfakes of Indian union ministers.
Watermarking as a limited compliance measure
The watermark cannot stop determined misuse: it only marks Anthropic's own output, while open-weight models like Moonshot's Kimi K3, Meta's Muse Glimmer and DeepSeek's V4 Flash can simply remove the marking code, leaving the systems most likely to be used for deception unreachable.
Key facts
- Company
- Anthropic
- Product
- Claude
- Technology
- Invisible watermark in text; C2PA signed metadata for generated files
- Legal basis
- Article 50 of the European Union's AI Act
- Effective date
- August 2, 2026
- Geographic scope
- Worldwide, not limited to Europe
- Not covered
- Open-weight models: Moonshot's Kimi K3, Meta's Muse Glimmer, DeepSeek's V4 Flash
- Related incident
- AI deepfakes of Indian union ministers circulated during the CJP protests and were debunked by the Press Information Bureau; one minister filed a police complaint











