Anthropic, the maker of Claude, has started embedding invisible machine-readable watermarks into everything its newest models generate, text and files, to comply with the EU AI Act's transparency rules. And it applied the change worldwide, not just in Europe. This is the content-marking side of the same Article 50 we covered last week, playing out in real life. It is proof the rules are real, that major labs are complying, and that the EU's standard is quietly becoming the global one. It also reveals a distinction every business publishing AI content needs to understand.
In Episode 2 of this series, we explained the rule that AI-generated content, especially deepfakes, must be labelled under Article 50 of the EU AI Act. That rule has two sides: a visible labelling duty on the people who publish content, and a machine-readable marking duty on the companies that build the AI. This month, we got a landmark example of the second side in action, and it is worth understanding, because it changes the landscape for anyone using these tools.
What Anthropic actually did
Anthropic confirmed that Claude models launched on or after August 2, 2026 now embed machine-readable marks into the content they generate. For text, this is an imperceptible watermark, a subtle statistical pattern woven into the writing that a reader cannot see but a detection tool can identify. Anthropic says it does not change the meaning, quality, or readability of the text, and that it survives copying, pasting, and light editing. For files like images, Claude attaches signed provenance metadata following the open C2PA standard, recording that the file passed through Claude and helping detect later tampering.
The company did this because it signed the EU AI Act's Code of Practice on the transparency of AI-generated content, a voluntary framework that gives signatories a presumption of meeting the Article 50 standard. In other words, a major lab looked at the rule we described last week and built compliance directly into its models.
Here is the detail that matters far beyond Europe. Anthropic applied this watermarking worldwide, to every user everywhere, not only to people in the EU. Nothing in the EU code strictly required a global rollout, but building one compliant system for everyone is simpler than running two. This is the Brussels Effect we described in our AI governance guide, no longer a theory: a European rule is now silently setting the standard for AI users on every continent. When the largest providers comply globally to satisfy Europe, the EU's rules effectively become everyone's rules.
The distinction every publisher must understand
This news makes one point from last week's episode sharper, and it is the single most important thing for your business to grasp. A machine-readable mark is not the same as a visible label. The watermark Anthropic embeds is a signal that a detection tool can read; it is not a notice that a human reader sees on the page. Under the AI Act, turning that invisible signal into a visible, human-understandable disclosure is a separate duty, and it largely falls on the organisation that publishes the content, not on the AI provider.
So this development does not let publishers off the hook. If anything, it clarifies the division of labour. The AI companies are handling the invisible, technical marking at the model level. But the visible labelling, telling your actual audience that a piece of content is AI-generated, remains your responsibility as the publisher. The watermark helps detection systems and platforms; it does not replace the clear on-screen label your audience is entitled to.
This is bigger than one company
Anthropic is not acting alone, and that is the real story. The C2PA provenance standard it uses for files has already been adopted by Adobe, the BBC, and several camera manufacturers to label the origin of images. OpenAI joined the same C2PA coalition earlier in 2026 and uses Google's SynthID system to watermark its generated images. What we are watching is not a single company's policy but the formation of an industry-wide norm: content provenance and AI marking are becoming a standard layer of how digital content works. For businesses, that means AI transparency is shifting from optional good practice to built-in infrastructure.
The honest limits
It is worth being straight about what these marks can and cannot do, because they are not magic. Anthropic itself acknowledges the limitations. File metadata can be stripped by converting the format, taking a screenshot, or re-saving the file. Heavy editing or translation of watermarked text can cause genuine AI output to go undetected. And the rollout drew real objections from users uncomfortable with their own edited or assisted writing being marked as AI-generated, and worried about false accusations. Marking is a useful signal, not a perfect one, and it does not settle the harder questions about what counts as AI-generated versus AI-assisted work.
What this means for you
The practical takeaway is simple. First, the rules in this series are real and are already being built into the tools you use every day, so treating them as a distant concern is a mistake. Second, the fact that AI providers now mark their content at the model level does not remove your own duty, as a publisher, to visibly label AI content for your audience where the rule requires it. And third, this is the direction of travel for the whole industry, so getting comfortable with AI transparency now is an investment, not a burden. Next week, we return to the numbered series with how to classify your AI system into the right risk tier.
Frequently asked questions
Does Claude now watermark everything it writes?
For models launched on or after August 2, 2026, yes. Anthropic embeds an invisible, machine-readable watermark into generated text and signed provenance metadata into generated files. The mark is not visible to a normal reader, and Anthropic says it does not change the text's meaning or readability. It applies worldwide, not just in the EU.
If Claude watermarks its output, do I still have to label AI content myself?
Yes, where the rule requires it. A machine-readable watermark is a hidden signal for detection tools, not a visible notice for your audience. Under the EU AI Act, the duty to provide a clear, human-visible label largely falls on the publisher of the content, not the AI provider. The watermark does not replace your labelling responsibility.
Why did Anthropic apply this worldwide instead of only in the EU?
Running one compliant system for all users is simpler than maintaining separate versions by region. This is the Brussels Effect in action: a European rule effectively becomes a global standard because major providers comply everywhere to satisfy Europe. It is a clear example of how EU AI regulation shapes the tools people use worldwide.
Can these watermarks be removed?
Sometimes. Anthropic acknowledges that file metadata can be stripped by format conversion, screenshots, or re-saving, and that heavy editing or translation of text can cause the watermark to go undetected. The marks are a useful signal rather than a tamper-proof guarantee, which is one reason regulators expect a multi-layered approach rather than relying on any single technique.
Is this just an Anthropic thing?
No, it is an industry shift. The C2PA provenance standard is also used by Adobe, the BBC, and camera makers, and OpenAI joined the same coalition and uses Google's SynthID to watermark images. Content provenance and AI marking are becoming a standard layer across the industry, driven in large part by the EU's transparency rules.
Follow the series, get compliant one rule at a time
This is a special update in our weekly guide to AI regulation, showing the rules in action at the biggest AI labs. Explore more clear, honest guides on AISetApp and follow the series each week.
Explore more on AISetApp- Anthropic Help Center, "How Claude marks AI-generated content," updated August 2026
- Euronews, TechCrunch, and The Next Web reporting on Anthropic's global watermarking and the C2PA standard, August 2026
- Forbes and Axios analyses of Claude text watermarking and its implications, August 2026
- EU AI Act (Regulation (EU) 2024/1689), Article 50(2) marking obligations and the Code of Practice on transparency
Reviewed August 2026. This is an explainer, not legal advice. The law is evolving; verify specifics with a qualified professional before acting.
Researched and drafted with AI assistance, reviewed and edited by Yasser El Hardouz, who takes editorial responsibility for this article.