What zero-width characters are, why scanners flag them as AI watermarks, and how to remove the residue without breaking emoji or Persian and Indic text.
The basics
A zero-width character is a real codepoint that renders with no width at all. Not a space you can see, not a layout gap, it takes up no room on screen yet sits in the text exactly like any other character would.
That is the specific trouble with them. A search, a diff tool or a slug generator reads every codepoint in a string, visible or not, so a zero-width character changes what a machine sees while leaving what a person sees untouched.
The characters
Worth telling apart rather than treating as one interchangeable problem.
Why scanners react
Chat interfaces use zero-width characters as internal bookkeeping: marking formatting boundaries, tracking cursor position through a copy, sometimes just an artefact of assembly before the text reached your clipboard. None of that requires a language model to explain, but AI chat interfaces are a common, repeatable source, which is why a scanner built for AI text checks these codepoints specifically.
A scanner flags a codepoint, not an author. It reports a specific invisible character at a specific position, a fact about the text. What produced the text is a separate question the character alone can't answer.
The honest part
Not everything flagged here is residue.
The fix
The goal is removing residue without removing characters doing real work.
Straight answer
No. These characters turn up in text from PDFs, word processors, translation software, and ordinary web pages with no language model anywhere near them. Finding one tells you something about formatting or copying, not about who or what wrote the words.
Treat a zero-width finding as a formatting fact worth cleaning up, not evidence about a document's origin.
Elsewhere
Deliberate uses outside AI text exist too: splitting a word inside a hashtag without visibly breaking it, slipping past a simple keyword filter, or making two identical-looking usernames register as different strings to a system comparing them byte for byte. None of that is about AI writing at all, and knowing this territory exists helps keep a single flagged character from being read as more significant than it is.
Aftercare
If the text has emoji, confirm compound ones still render as a single glyph rather than separate pictures side by side. If it's in Persian or an Indic script, read it back and confirm the letterforms still look right. A cleaning pass that trades an invisible problem for a visible one hasn't actually helped, whatever it did to the byte count.
A character, U+200D, instructing two adjacent characters to render as one joined glyph, most visibly used in compound emoji.
A character, U+200C, preventing two adjacent letters from joining into a connected glyph, with a real role in Persian and some Indic scripts.
Four: zero-width space, zero-width non-joiner, zero-width joiner, and the word joiner, which behaves the same way under a different name.
It can, if done blindly. A joiner gluing an emoji sequence together should be recognised and left alone.
Not functionally. In English text it's almost always residue.
Yes, to separate a prefix or suffix from its stem so both display correctly without visually joining.
Several do, to stop two consonants from merging into a single conjunct glyph.
No. The same characters appear in PDFs, translation tools and word processors with no model involved.
The formatting and copy mechanism is similar across chat tools, so text from any of them can carry the same characters.
Scan it with a tool that names each character and position first, rather than running a blind delete-all pass.
It shouldn't, if the cleaner recognises which non-joiners are doing linguistic work. Review flags in that kind of text before a bulk removal.
No, separate categories with their own causes. This guide is specifically the zero-width and word-joiner family.
Yes, they've been used to split a word without visibly breaking it, or to slip past a simple filter.
Character count drops. Word count is unaffected, since none of these four are ordinary spaces.
No, this is a guide. The watermark remover and checker on the homepage does the actual cleaning, with the same emoji and script-aware handling described here.
Not yet. Paste, scan, clean, repeat for the next document.
Yes, these characters are a property of the text itself, not of one chat interface.
Yes, it reads the same way on a phone.
No. There is no account system anywhere in this app.
Yes. That guide covers all 14 hidden characters this app checks for. This one narrows in on the four zero-width ones specifically, in more depth.