Anthropic will watermark text from Claude to meet EU law
Every piece of text Claude generates will carry a hidden label so other systems can identify it as AI-made. The change is mandatory under European regulations that took effect this August.
Key points
- Anthropic confirmed on an updated support page that all Claude models released after 2 August 2026 will automatically watermark AI-generated text and files.
- The EU AI Act's Transparency Code, which came into force on 2 August 2026, requires AI companies to mark AI-generated content so other software can detect it.
- For files, Anthropic is using C2PA, an open technical standard for labelling digital content.
- The watermark travels with copied and pasted text, and may survive some editing.
- Anthropic, Google, Meta, Microsoft, OpenAI, Black Forest Labs, Synthesia and others have all committed to following the EU code.
Anthropically, the company behind the Claude family of AI assistants, has confirmed it will watermark every piece of text its models produce. A watermark here isn't a visible stamp like the ones on stock photos. It's a hidden signal embedded inside the text itself that software can detect, even if a human reader cannot see it.
The EU AI Act's Transparency Code, a set of rules for companies that make AI tools, came into force on 2 August 2026 and requires AI companies to mark content their systems generate so other systems can identify it. We covered the rules' arrival in our 31 July story on mandatory labels for AI-generated images, audio and text. First reported by TechCrunch AI, Anthropic's confirmation that it's complying affects Claude across every surface it runs on.
What exactly changes for people who use Claude?
Every output from a Claude model released after 2 August 2026 will carry the watermark automatically, with older models getting the same treatment over time.
The watermark is baked into the text, not tacked on separately. Copy a paragraph Claude wrote and paste it into an email or a document, and the label comes with it. Anthropic's support page notes it "may persist through some editing", though the company hasn't said how much editing strips it away.
The watermark applies to Claude.ai, the Claude API (the programming interface developers use to build products on top of Claude), Claude Code, Claude Cowork and Claude Tag.
For files rather than plain text, Anthropic is using the C2PA standard. C2PA, short for Coalition for Content Provenance and Authenticity, is an open technical framework that embeds origin information directly into a file, much like a digital birth certificate attached to an image or document.
Should ordinary users care about this?
For most people using Claude day-to-day, nothing about the experience changes. The watermark's invisible and automatic.
Where it matters is downstream. A teacher checking an essay, a platform screening for AI-generated spam, or a journalist verifying a source could all use detection tools that read these labels. The watermark doesn't guarantee detection catches everything, but it gives those tools something real to look for.
Anthropically isn't alone. Google, Meta, Microsoft, OpenAI, Black Forest Labs and Synthesia have signed up to the EU code. AI music platform Suno also announced watermarking last week after a run of legal challenges. The honest read on all this: watermarking is becoming a floor, not a ceiling, and the companies that move slowly on it are taking a regulatory bet they may not win.
What to watch for: If you use Claude-generated text in professional or published work, be aware the text carries a machine-readable label. Some platforms may flag or filter it automatically.
