Watermarking: how AI vendors propose to mark their own text
Text watermarking is a technique in which a language model vendor nudges its own model's word choices in a secret, statistical pattern so that the vendor can later test a passage and say "this came from us". It is different from the watermark on a photo: nothing is visible, and the text reads normally. Humanize 360 does not watermark anything and does not insert hidden characters, homoglyphs or zero-width tricks anywhere in its products. This page explains the idea, because people confuse the two.
The general scheme, as published by researchers and described by vendors, works like this. At each step the model has a list of candidate next words. The watermark secretly splits that list into a preferred group and a non-preferred group, using a key, and leans slightly towards the preferred group. Any single sentence looks normal. Over hundreds of words the preferred group appears more often than chance would allow, and someone holding the key can measure that.
What a watermark is not
- It is not invisible Unicode characters between letters. Those are a crude trick, easily stripped, and detectors flag them.
- It is not homoglyphs, where a Latin "a" is swapped for a Cyrillic one. Screen readers choke on those and so do spell checkers.
- It is not metadata in a file. The pattern lives in the words themselves.
- It is not something a reader can see. Without the key, a watermarked paragraph is just a paragraph.
Why watermarks are hard to rely on
The vendor has to choose to apply it, and only that vendor can check it. Text from a model without a watermark shows nothing. Short passages do not contain enough words for the statistics to settle. Editing, paraphrasing or translating a passage erodes the pattern. And a watermark says which model wrote something, not whether using that model was allowed. Public discussion among vendors has gone back and forth on whether to deploy it widely, and we do not know the current internal state of any of them.
Where Humanize 360 stands
Our engine is our own code, not a language model, so there is nothing to watermark and no vendor key to worry about. We never add hidden characters to output, in either direction: we do not insert them, and the Changes view shows every edit in plain text so you can see there is nothing else there. What the detector measures is the ordinary statistical texture of writing, which any reader can also feel.
A worked illustration
Suppose the secret rule preferred words starting with letters from the first half of the alphabet. "The delivery arrived before noon" fits it four times out of five. "The package showed up before twelve" fits it two times out of six. One sentence proves nothing. A thousand words leaning the first way, consistently, would be a fingerprint. Real schemes are far subtler than this, but the logic is the same.
Common questions
Does ChatGPT watermark its text?
We do not know the current state of any vendor's deployment. Research on the method has been published and vendors have discussed it publicly, but whether and where it is switched on is not something an outside tool can verify.
Can a detector see a watermark?
Only the party holding the key can test for it. Third-party detectors, including ours, measure ordinary statistical texture instead, which works on any text regardless of source.
Does Humanize 360 add anything hidden to my text?
No. No zero-width characters, no homoglyphs, no invisible markup. Detectors strip those and flag them, and they break accessibility tools. Every change we make is visible in the Changes view.
Does humanizing remove a watermark?
Any substantial rewrite disturbs a statistical watermark, but we make no claim about it because we cannot test for keys we do not hold. Our purpose is text that reads human, not text that hides its origin.