Zero-width space, U+200B

A character with no width and no visible form. It occupies a position in the string, counts towards its length, and shows nothing at all. It is the character most people mean when they say an AI put something invisible in their text.

How it works

In UTF-8 the zero-width space is three bytes, e2 80 8b. A text renderer looks up its properties, finds a character with the general category Cf, format, and a width of zero, and advances the cursor by nothing at all. It is a real character occupying a real position: it counts in a string's length, it moves every index after it by one, and it round-trips through the clipboard, through JSON, through a database column and through a diff, because every one of those layers is copying bytes rather than interpreting glyphs. That combination, invisible on screen and fully present in the data, is what makes it break things silently. A comparison is a byte comparison, so two strings that read identically are unequal. A tokeniser splitting on whitespace does not treat it as whitespace. A compiler reading an identifier includes it in the name. Detection is trivial once you look for it, which is why AI detectors do: it is a single code point with no legitimate reason to appear mid-word in English prose.

What it is legitimately for

It marks a place where a line may break without adding a visible space, which matters in scripts and in long unbroken strings like URLs. Used deliberately, it is a typesetting tool.

What it breaks

  • An identifier containing one compiles as a different symbol, so code fails with an error naming a variable that looks identical to the one you defined.
  • A LIKE or equality query silently misses rows whose text looks the same on screen.
  • A git diff reports a line as changed when nothing visible changed.
  • Deduplication and comparison treat two identical-looking strings as different.

How to find it yourself

Pipe the text through a hex dump and look for e2 80 8b, or search with a regex that names the code point rather than the character, such as grep -P for the escape form. Most editors can be told to reveal invisible characters.

What Clipboard Sanitizer does

Removed everywhere. It has no visible form, so removing it cannot change how your text reads.

Questions

Is a zero-width space a watermark?

No vendor has ever said so. It is best understood as an artifact of how models are trained on text that legitimately contains these characters. That said, a detector does not care why it is there.