The watermark travels with the text when you copy and paste it. That’s the detail worth paying attention to in Anthropic’s latest compliance update, because it’s more ambitious than it sounds. According to TechCrunch, Anthropic confirmed it will watermark text and files generated by Claude models, including outputs from the API, Claude Code, Claude Cowork, and Claude Tag. The trigger is regulatory: the EU AI Act’s Transparency Code took effect on August 2, requiring AI companies to mark generated or edited content in a machine-readable way.
All Claude models released after August 2 will carry watermarking automatically. For files, Anthropic is using C2PA, an open standard backed by a broad industry coalition. For text, the watermark is embedded directly in the output itself, meaning it can survive a copy-paste and may persist through light editing. Anthropic says support will extend to older models too, though timing on that hasn’t been specified.
What remains unclear is exactly how much editing strips the watermark out. That’s not a minor detail. If a few sentence rewrites are enough to remove it, the practical value drops significantly. Anthropic hasn’t clarified this yet, which is a gap worth watching for developers building on the API who may face compliance questions from enterprise customers operating in the EU.
The broader context here matters. Anthropic is not moving alone. Black Forest Labs, Google, Meta, Microsoft, OpenAI, and Synthesia have all committed to the EU’s code. Last week, AI music platform Suno said it would mark AI-generated tracks following legal pressure. Substack partnered with Pangram Labs last month to flag AI-written newsletter content, with CEO Chris Best publicly calling out what he termed “Claudefishing,” where writers use Claude to generate content and publish it without disclosure.
So this is less a story about Anthropic specifically and more about a compliance wave hitting the AI industry at the same time. The EU’s framework is functioning as intended: it’s getting companies to act in ways market pressure alone didn’t. Whether these watermarks hold up under real-world use is a separate question, and one that regulators, researchers, and developers will be stress-testing over the next year.
For founders and developers shipping products in or into Europe, the immediate takeaway is practical. If your product is built on Claude or any of the named platforms, watermarking is now on by default. You don’t need to configure it, but you do need to understand what it means for your users and your compliance posture. That’s a conversation worth having with your legal team sooner rather than later.
