“The watermark lives in the words Claude picks, one at a time.”

Anthropic announced watermarks for Claude-generated text. As someone who has spent years embedding metadata into photography, I had some thoughts.

From the research I had done, I thought it was adding metadata onto a file, much as you do with photos. I totally underestimated it, how it actually works is much more interesting.

Not a file header you can strip

Turns out there's no header, no field you can open up and strip out, the way there is with a photo. The watermark lives in the words Claude picks, one at a time, nudged just enough by a hidden key that the sentence still reads completely normal. Google does the same thing over on Gemini, it's called SynthID, and the trick holds up because you're never looking at one word, you're checking the pattern across hundreds of them.

Where it breaks down

It has real limits, though. Ask for the capital of France and there's only one word that fits, so there's nowhere to hide a signal. Translate the text, or run it through a paraphraser, and the words change enough that the mark washes right off.

Why now: the EU AI Act

A lot of why the labs are doing this at all comes down to the EU, their AI Act now requires generative text to carry a machine-readable mark. That's a completely different animal from tools like Turnitin or Pangram, which have no key to check against, they're just guessing at your writing style, and their false-positive record is rough enough that it's already gotten real academic papers rejected.

I'd like to thank Squintist for the YouTube video that provided a deep dive into how this works.

If you want to dig into how a specific AI feature you're evaluating actually works under the hood, that's exactly the kind of conversation worth having. Find me through the contact section of this site.