Anthropic started watermarking text on August 2. Anything written by Claude models from that date forward carries a mark, and older models are being retrofitted. It applies across the API, claude.ai, Claude Code, Cowork and Claude Tag, including Claude accessed through AWS, Google Cloud and Microsoft Foundry. Images and SVGs get signed provenance metadata using the C2PA standard instead.
When a supported model generates text, it "weaves an imperceptible watermark directly into the text itself," and that mark "will travel with the text when it's copied and pasted elsewhere, and may persist through some editing." Anthropic says the reason they're doing this is transparency, under a code of practice tied to Article 50 of the EU AI Act.
You can't read the mark, but they can
Anthropic says it is "working to enable users and other third parties to detect Claude's embedded watermarks and provenance metadata," with technical documentation to follow. So the detector isn't out yet but it's coming.
But the important thing is that Anthropic can read it right now. It's their signal, in their models, using their key. The marking is worldwide, not limited to the EU where the rule applies. A compliance requirement covering European users produced a global rollout.
What a watermark is good for
The single biggest technical problem facing anyone training a frontier model is that the internet is filling up with AI slop and machine text, and models trained on machine text get worse. That's model collapse, and it's documented rather than theoretical. The scale of contamination is not small either. A scan of over a million social posts found more than 40 percent of LinkedIn longform reading as fully AI-generated, and about a quarter of all long posts across platforms.
So every AI lab now needs to solve the same problem: scraping the open web, which parts were written by a machine, so we can drop them and which ones are by humans so we can consume that data. An embedded watermark is something cheap, deterministic, and machine-readable at exactly the scale you need.
A watermark that can survive copy-paste and some editing is close to a perfect filter for training data.
Where the theory breaks down
Anthropic's watermark only marks Claude output. It does nothing about text from GPT, Gemini, Llama, Kimi or the dozens of open models people run locally, and that's most of the contamination. As a defense against model collapse, marking only your own output is like sandbagging one house on the street.
The labs also don't need a watermark to know what they generated. They have logs. If the goal were only to exclude your own output from your own training set, you'd solve it internally without embedding anything in the text at all.
What I think the real motive is
No single company gets much benefit from watermarking only its own output. Anthropic can already identify plenty of Claude-generated content through its own systems, and a Claude watermark does nothing to flag text written by GPT, Gemini, Kimi, Llama or anything else.
The equation changes if every major AI company starts doing it. If Claude, GPT, Gemini and the rest all embed machine-readable marks into what they generate, the industry suddenly has a common way to identify synthetic text across the open web. A company scraping billions of pages for its next training run would no longer need to guess which passages were written by a model. It could detect the mark and leave them out.
That is why I keep wondering whether cleaner training data is the bigger reason AI companies are embracing watermarks.
The official explanation is transparency, and I think that explanation is just a part of it. The EU AI Act is pushing companies in exactly this direction. But the same infrastructure also solves one of the industry's nastiest long-term problems: models increasingly training on the output of other models.
And unlike a public-facing transparency tool, that benefit starts paying off immediately for the labs. Users still cannot check Claude text for Anthropic's watermark themselves. Anthropic can. The mark is already being embedded, while the detector for teachers, editors, employers and everyone else is still something the company says is coming.
If that order eventually reverses and Anthropic releases a genuinely useful public detector, the transparency argument gets stronger. If the watermark remains far easier for AI companies to use at scale than for ordinary people to check, the training-data explanation starts looking a lot harder to dismiss.
Because once all the major labs are marking their output, they have effectively built a machine-readable layer across the internet that says: this text came from AI, skip it when you train the next model. That is probably the main reason for the watermarks.
Sources
Anthropic: How Claude marks AI-generated content - The embedded text watermark and C2PA provenance metadata for files, the August 2, 2026 model cutoff and retrofitting of older models, coverage across the API, claude.ai, Claude Code, Cowork and Claude Tag, persistence through copy-paste and some editing, the statement that detection for users and third parties is still being worked on, the Article 50(2) code of practice, and the caveat that a detected mark signals processing rather than authorship.
TechCrunch: Anthropic says it will watermark text generated by its AI models - Independent coverage of the announcement and its scope.
Interesting Engineering: Anthropic puts hidden watermarks on Claude text under new EU rules - The worldwide rollout beyond the EU, coverage of Claude accessed through AWS, Google Cloud and Microsoft Foundry, and the limitations that heavy editing weakens the mark and short passages may not carry enough signal to detect.
