On August 11, Anthropic announced that Claude would start watermarking the text it generates. The reaction on X was immediate: dozens of users said they were canceling their subscription. The frustration wasn't really about what the watermark does. It was about what people assumed it did.
That's worth untangling, because Claude's watermark isn't what most people think it is.
What it actually is
The watermark is a statistical pattern woven into the text at the moment Claude generates it, not something added afterward. The mechanism, based on the SynthID Text approach Google DeepMind published in 2024, works on low-stakes word choices: when the model has to pick between two options that convey the same thing, like "overcast" versus "grey," it uses a key along with the preceding words to decide which one to pick. That choice doesn't change the meaning or quality of the response. It adds no extra tokens. It costs nothing extra. And it isn't visible to the reader. It can only be detected with the key Anthropic controls.
It also doesn't identify a specific person, organization, or conversation. The mark says Claude generated the text. It doesn't say who asked for it.
Why it exists now
Every Claude model released on or after August 2, 2026 applies this watermark by default, worldwide, with no opt-out. The date isn't a coincidence. That same day, the EU AI Act's Transparency Code took effect, requiring AI-generated content to be marked in a way other systems can recognize. Anthropic signed that code in July alongside roughly 190 other signatories. The watermark is, first and foremost, a regulatory response. Applying it globally rather than just in the EU is the part Anthropic chose on its own.
What breaks it, and what doesn't
Light editing probably won't remove the watermark. A complete rewrite, where every word gets replaced, will. That leaves an uncomfortable gray zone: if you rewrite enough to erase the mark, at what point did the text stop being AI-generated. Anthropic doesn't answer that question, and it probably isn't theirs to answer.
Short text is also a problem for the system. The watermark needs enough words with real choice behind them to leave a reliable signal. A block of code, or a purely factual passage with almost no word-choice flexibility, carries a much weaker mark, close to none in the case of code.
| Factor | Effect on the watermark |
|---|---|
| Light editing or proofreading | Probably doesn't remove it |
| Complete rewrite (every word replaced) | Removes it |
| Short passages | Unreliable signal |
| Factual content with little word variation | Weaker signal |
| Code | Negligible signal |
What it means for anyone writing with AI
For someone who uses Claude to draft something and then edits it heavily, this changes little. The watermark survives light editing, but not text that's already been through real editorial judgment. For someone who copies and pastes the response as-is, the mark travels with the text, even if Claude only fixed the spelling.
Anthropic plans to release a detection API, though it hasn't given a date or implementation details yet. That's different from the AI detection tools that already exist, which look for writing patterns. The watermark doesn't look for patterns. It verifies a key. Those are two different ways of trying to answer the same question, and neither can confirm whether a human wrote something. They can only say, with varying degrees of certainty, whether Claude was involved.
For anyone producing content for growth or marketing, the question that actually matters isn't whether Google or a client will detect that you used AI. It's whether the editorial judgment applied before publishing is enough for the result to be good, watermark or not. The watermark identifies model involvement. It doesn't identify quality. That's still a human call.
The watermark doesn't settle who wrote what. It just makes it harder to pretend no one used AI to do it.