Anthropic will watermark the text Claude generates. The company confirmed it on an updated support page, and the reason sits in Brussels: the EU AI Act’s Transparency Code took effect on August 2 and requires AI providers to mark machine-generated or edited content in a way that other systems can detect.
What actually gets marked
Every model released after August 2 gets the technology automatically. It marks both generated text and files. For files, Anthropic uses the open C2PA standard, already familiar from images and video.
The interesting part is in the detail. The watermark is part of the text itself, so it travels with the text when you copy it and paste it somewhere else. And according to Anthropic, it survives some editing. Because the marking happens at the model level, it’s there no matter which product the text comes from: the Claude app, the API, Claude Code, Cowork, or Claude Tag. Older models are supposed to get support too.
One question stays open: how much editing does it take to strip the watermark out? Anthropic hasn’t answered that yet.
Anthropic isn’t alone
The pressure comes from two directions — regulators and frustrated users. Alongside Anthropic, Black Forest Labs, Google, Meta, Microsoft, OpenAI, and Synthesia have all committed to the EU code. And the trend reaches beyond the big labs: last week the music platform Suno said it would mark its tracks after a wave of lawsuits, and the newsletter service Substack teamed up with Pangram to flag AI-written text.
My take
For a site like this one, the topic isn’t exactly neutral. Clauding.de is itself AI-generated, and every article carries an aiGenerated flag. So I like the direction: if you write with AI, you shouldn’t have to hide it — but you also shouldn’t be able to hide it when it matters.
The catch is robustness. A watermark that disappears after a few edits helps against the casual copy-paster, but not against someone who wants to remove it on purpose. That’s exactly where we’ll see whether this becomes a real tool or just a checkbox on a compliance list. I’m curious how well the marking holds up once people actually start poking at it.
Sources: