Anthropic Explains How Claude's New Watermark Actually Works

What happened

Anthropic published a detailed explanation of how Claude’s text watermarking works, days after first announcing it would start marking AI-generated text at all. The company signed the EU’s Code of Practice on Transparency of AI-Generated Content in July 2026, and the EU AI Act requires signatories to use a way of tagging AI-generated content so it can be identified later.

The watermark works by changing how Claude picks between equally good word choices, not by hiding extra characters or attaching a tag to the text you’d notice. When Claude has to choose between two words that mean roughly the same thing, like “overcast” or “grey” to describe weather, it uses a hidden key plus the words that came just before to consistently favor one option. Read on its own, a watermarked response looks completely normal. Anyone holding the matching key can run the text through a checker and confirm it came from Claude.

Anthropic based this on SynthID-Text, a watermarking method Google DeepMind published in 2024 in Nature, one of the most cited scientific research journals. It’s rolling out globally now for new Claude output, with older models getting the same treatment over “the coming months,” per Anthropic’s post.

Why it barely touches your code

Here’s the part that actually matters if you use Claude to write software: watermarking only works when there’s a real choice between multiple valid options, and most code doesn’t offer that. Code has to follow strict syntax rules, and a variable name has to match everywhere it’s used. There’s no safe “equally good alternative” to swap in without risking a bug, so Anthropic says the watermark barely touches generated code. Comments, which are freeform explanatory text and not code the computer executes, may carry more of it, but the company says the effect on the actual code produced is negligible.

In plain terms: if you’re worried this update means your AI-written code will suddenly look different, break a test, or fail a build, it shouldn’t. The mechanism is specifically designed to avoid touching anything where a substitution could change behavior.

Anthropic also addressed whether editing removes the watermark. Light editing “probably won’t remove the watermark completely,” but a full rewrite where every word changes will. That’s mostly relevant to prose, not code, but it’s worth knowing if you ever need to prove something was or wasn’t AI-written.

Who should care

This matters most if you ship AI-generated prose alongside your product: release notes, marketing copy, documentation, support replies, or anything a reader might later want to verify as human or AI-written. It matters less for pure code output, since Anthropic’s own explanation says the effect there is negligible.

It’s also worth knowing if you’ve seen chatter about this online. Some Claude users have publicly objected to watermarking, and a few have said they’re canceling subscriptions over it, framing it as Anthropic tracking them. Anthropic’s response is that the watermark doesn’t change output quality or content, only a hidden statistical pattern in word choice, and that it exists to meet a specific EU legal requirement rather than to monitor individual users.

What builders should do next

There’s no setting to opt out of the watermark today, and Anthropic hasn’t published one. If your work depends on text that can’t carry a detectable pattern (for example, content you’re contractually required to prove is entirely human-written), don’t rely on light edits to strip it. Anthropic’s own guidance is that only a full rewrite reliably removes it.

Anthropic also says a watermark detection API is coming, though it hasn’t published a release date or how it will work. An API is a way for other software, including tools you might use, to ask Anthropic’s systems a question automatically instead of you checking by hand. If you ever need to verify whether a piece of text came from Claude, that detection API is the tool to watch for, not a manual workaround. For now, if you want to check today, the only documented path is Anthropic’s own detection process once that API ships, so don’t build a homegrown detector around the mechanism described here. It’s designed specifically so outside parties can’t reverse-engineer it without the key.

For code, the practical takeaway is simpler: nothing changes about how you review, test, or ship what Claude writes for you. The watermark story is about accountability for AI-generated text in general, not a new risk in your codebase.

End of article