Anthropic confirms Claude text gets permanent watermarks to satisfy EU AI Act

Anthropic just confirmed it’s baking digital watermarks directly into the text generated by Claude and its other models. The move is a direct response to the EU AI Act’s Transparency Code, which kicked in on August 2 and demands that AI-generated or edited content be machine-readable. If you’re using a model released after that date, the watermarking is already automatic—no opt-out. And it’s not just a metadata tag. The company says the signal is embedded at the model level, meaning it lives inside the text itself. As their support page puts it: “Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing.”

That last part is the real kicker. Most watermarking schemes fall apart the moment you hit Ctrl+C. Anthropic is claiming persistence, even through light edits. The exact robustness remains a black box—they haven’t clarified how much tweaking it takes to scrub the signal clean. We’ve asked. It’s the same opacity problem that plagued earlier tools like OpenAI’s classifier, which was shelved partly because it was too easy to fool. For files, Anthropic is using the C2PA open standard, a more established approach that acts like a nutrition label for digital content. That’s the safer, more proven bet here.

The compliance scramble isn’t just an Anthropic problem. Substack recently partnered with Pangram to sniff out AI-generated slop, with CEO Chris Best coining the term “Claudefishing” to describe people who pass off AI text as their own. Music platform Suno is watermarking tracks after a wave of legal battles. And the EU’s list of companies promising to play ball reads like a who’s-who of AI: Google, Meta, Microsoft, OpenAI, Black Forest Labs, and Synthesia are all on the hook. Nobody wants to be the test case for EU fines.

What’s genuinely new here is the scale. Anthropic isn’t just watermarking one product—this applies across the Claude platform API, Claude Code, Claude Cowork, and Claude Tag. That ubiquity could actually shift norms, especially in enterprise settings where compliance teams are terrified of unlabeled AI output slipping into customer-facing work. The lingering question is whether persistent watermarking changes user behavior or just drives people to older, unmarked models. Anthropic says it’s extending support backward to older releases, but the timeline is vague. If the watermark truly resists editing, content moderation might finally have a technical enforcement layer—not just a policy one.