By Shannon Glennon, AI Technology Transformation Specialist, Syracuse University ITS
If you’ve heard that Claude puts a “watermark” on its writing, you probably pictured something you’d have to search for and delete, like a logo behind a stock photo, or like the watermark added to videos in TikTok or CapCut. Here’s the twist: there’s nothing to find. The watermark is a pattern in the words themselves. Here’s the full story.
So what is it, exactly?
When Claude writes, it picks one word at a time, and typically several words would work just as well. “The meeting was productive” and “The meeting was useful” both get the job done. The watermark gently tilts those picks in a set direction. One tilted pick means nothing. Across a few paragraphs, the tilt adds up to a pattern. Nothing is added to the text and there are no hidden characters. Anthropic says the method is a version of SynthID-Text, which Google DeepMind published in a 2024 Nature paper.
What’s the “secret key”?
A secret digital key decides which words get tilted. Think of it as a code only Anthropic knows. Without the code, the pattern just looks like normal writing. With it, someone can run a count and see whether the word choices lean the way Claude’s would.
Why is it there?
Anthropic says it’s part of complying with Article 50 of the EU AI Act, which is about making AI-generated content identifiable.
Which Claude models use it?
Every Claude model launched on or after Aug. 2, 2026 uses it, which includes the three at the top of your model menu: Opus 5.5, Sonnet 5.5 and Haiku 5.5. Older models have been joining since then. It applies in Claude, Claude Code and Cowork.
Can anyone check for it?
Not by reading the text, and not with a free tool. The detection tool is in private preview for eligible organizations like regulators, media outlets and researchers. There’s no public checker currently.
Does it stick around?
Yes. The mark travels with the text when you copy and paste it, because it lives in the word choices, not in a file tag. Light editing probably won’t remove it, while replacing every word will. If you ask Claude to proofread your own writing, there’s very little for the watermark to attach to, since nearly all the words are yours.
Does it prove Claude wrote something?
No. Anthropic says a detected mark doesn’t prove authorship, and it contains no information about who used Claude. It also can’t tell the difference between “Claude wrote this” and “Claude (heavily) edited this.” It’s a clue, not a verdict.
What about images?
Images get a label instead. PNG and JPEG files from Claude carry Content Credentials, a metadata label anyone can read with the free Claude Content Checker.