Anthropic Watermarks Claude Text and Files Under EU AI Act
Anthropic has signed the EU AI Act's Article 50(2) transparency code and will embed invisible watermarks in Claude text and C2PA metadata in generated files, worldwide, starting with models launched August 2, 2026.
Key takeaways
- Anthropic will embed hidden watermarks in text from new Claude models and signed C2PA metadata in generated .svg, .png, and .jpg files, applied worldwide.
- The move fulfills Anthropic's signature on the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content.
- Models launched in the EU on or after August 2, 2026 support the marking at launch; earlier models are covered by a transition period.
- Two of Article 50's four exemptions apply to edited copy, meaning a Claude watermark can turn up on text a publisher has no legal duty to label.
- About 190 signatories had joined the EU code by the end of July, per the European Commission, including Google, Meta, Microsoft, Mistral, and OpenAI.
What Happened
Anthropic will start embedding invisible watermarks in text generated by new Claude models and attach signed provenance metadata to generated image files, a rollout that applies worldwide rather than just inside the EU. Search Engine Journal reported on August 11 that the marking fulfills Anthropic's signature on the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, which it joined as a provider of both generative models and generative systems.
The watermark covers Claude output from the API, the Claude apps, Claude Code, Cowork, and Tag, plus access through AWS, Google Cloud, and Microsoft Foundry. Anthropic says the mark doesn't alter meaning, quality, or readability, survives copy-paste, and "may persist through some editing." Generated .svg, .png, and .jpg files carry signed metadata under the open C2PA standard, recording how a file was made and whether it's been altered.
Signatories to the EU AI Act's Article 50(2) transparency code by end of July, per the European Commission
European Commission, via Search Engine Journal, 2026
Why It Matters For Marketers
A detected mark only means Claude may have processed the content, not that Claude authored it. Anthropic's own limitations section warns that proofreading, translation, summarizing, and file conversion can leave a mark even when the ideas and text originate elsewhere, which means AI-assisted human copy can get flagged the same as fully generated drafts.
Anthropic isn't alone. OpenAI shelved its own watermarking plan in September 2024 after finding that nearly 30% of ChatGPT users said they'd use the product less if it were watermarked, a detail that shows how sensitive this feature is for adoption. It's a tension we've also flagged in coverage of the EU AI Act's broader labeling rule, which takes effect this month.
What To Do This Week
- Check your editorial sign-off process against the Article 50 exemptions during your next martech stack audit to confirm which Claude-assisted content genuinely needs disclosure.
- Detection tools for the watermark are still unpublished, so don't promise clients or legal teams a way to verify marks yet.
- Flag any generated .svg, .png, or .jpg assets in your DAM for the new C2PA metadata before assuming files are unmarked.
Get more AI governance news and compliance breakdowns for marketers.
Advertiser disclosure: some links in our articles are affiliate links, and CMO Mag may earn a commission or referral fee if you sign up or buy through them, at no cost to you. It never affects our editorial coverage. See our advertising & affiliate policy.
More in AI in Marketing
View all
EU AI Act's AI Content Labeling Rule Sparks SEO Debate
A widely shared Reddit thread is drawing fresh attention to the EU AI Act's transparency requirements for synthetic content, and marketers running EU campaigns need to know where their obligations start.
AI Overviews Cite Trust Signals, Not Just Rankings
A live webinar detailed by Search Engine Journal shows Freshpet holding the top organic spot while AI Overviews still name competitors instead, and lays out the verifiable trust signals that changed that.
Google Folds Top Stories Into AI Overviews, robots.txt Can't Stop It
New tracking from NewzDash shows Google is folding entire Top Stories carousels into AI Overviews on trending news queries, and the robots.txt block publishers reach for first has no effect on it.
Discussion
No comments yet. Be the first to say something worth reading.