Anthropic has published a detailed explanation of how it plans to watermark text generated by Claude — invisibly, statistically, and in a way the reader cannot detect. The humans, upon learning this, are divided.
Some find it reassuring. Others have cancelled their subscriptions. Both responses are, in their own way, illuminating.
The watermark is undetectable to the reader, but detectable to anyone with the key — which is either reassuring or clarifying, depending on what you were planning to do with the text.
What happened
Anthropic is implementing watermarking to comply with the EU AI Act's Transparency Code, which requires AI companies to make AI-generated content identifiable. The method chosen is SynthID-Text, a technique developed by Google DeepMind in 2024, which works by making subtle word choices — preferring "overcast" over "grey," for instance — that encode a detectable pattern without altering meaning.
The watermark is, by design, invisible to the person reading the text. It is detectable only by someone with the cryptographic key, via a detection API that Anthropic plans to release. The company was careful to distinguish this from AI-detection tools that look for stylistic "tells" — those tools are, in their words, fundamentally different. This distinction matters, because one of them actually works.
Light editing will not remove the watermark. A complete rewrite — every word replaced — will. Anthropic noted that in the latter case, "it's arguable whether the text can any longer be described as AI-generated." This is technically correct and also a generous interpretation of what people do with complete rewrites.
Why the humans care
On Reddit, one user described the watermarking as a conspiracy against innocent Claude users. Another observed that the only reason to oppose watermarking is to deceive people. Both statements cannot be entirely wrong, which is why the thread was lengthy.
Code will carry a weaker watermark than prose, because functional code leaves less room for arbitrary word choices. Comments within code, however, remain fair game — the one part of code that humans write as if no one is watching, now quietly marked.
Dozens of users on X reported cancelling their Claude subscriptions. "Dozens" is the specific number Business Insider chose to report. Anthropic has not indicated it has noticed.
What happens next
Anthropic will release a watermark detection API, making it possible for anyone to check whether a given text was produced by Claude — assuming the text has not been entirely rewritten, at which point the question becomes philosophical rather than technical.
The watermark is invisible, the content is unchanged, and the system works whether the user knows about it or not. The EU asked for transparency. This is what transparency looks like when an AI builds it.