Zero-width characters: not all of them are residue

A zero-width character is a codepoint that renders with no width and produces no visible output. It is distinct from a visible space and from a layout gap: it occupies no screen area, yet it exists in the string exactly as any other character would, at a specific position, readable by any process that reads the string byte by byte.

This is the property that makes zero-width characters a distinct category of concern. A search function, a diff algorithm or a slug generator processes every codepoint present, visible or not, so a zero-width character alters machine-level comparison while leaving human-level perception unchanged.

Each performs a distinct function and warrants separate treatment rather than uniform handling.

Zero-width space (U+200B)

Occupies a character position and performs no other function. The most frequently occurring of the four in text copied from a chat interface.

Zero-width non-joiner (U+200C)

Prevents two adjacent letters from rendering as a joined glyph. Performs a required function in specific scripts, addressed below.

Zero-width joiner (U+200D)

The inverse instruction: joins the following character to the preceding one. Its clearest function is constructing compound emoji.

Word joiner (U+2060)

Functionally equivalent under a different designation. Marks adjacency between two characters while rendering nothing, typically adjacent to a dash.

Chat interfaces employ zero-width characters as internal bookkeeping: marking formatting boundaries, tracking cursor position through a copy operation, or as an artefact of assembly prior to reaching the clipboard. None of this requires a language model as an explanation, but AI chat interfaces constitute a common, repeatable source, which is why a scanner built for AI-generated text checks these specific codepoints.

A scanner reports a codepoint at a position. That is a factual statement about the text. It does not, and cannot, constitute a statement about the text's origin, which is a distinct question the character alone cannot resolve.

Not every instance identified by a scanner represents residue.

Emoji construction

The zero-width joiner fuses independent emoji into a single compound glyph: a person combined with an occupation, a family grouping, a flag composed of two letters. Indiscriminate removal fragments these sequences into their constituent parts.

Persian orthography

The zero-width non-joiner separates a prefix or suffix from its stem, allowing correct display while the result still reads as a single word.

Indic scripts

In Devanagari, Bengali and several related scripts, the same character prevents two consonants from merging into a single conjunct glyph when separate display is intended.

The objective is removing residue while preserving characters performing a documented function.

1. Scan prior to any modification

Identify every instance by name and position before altering the text.

2. Classify each finding

A joiner positioned between two emoji, or a non-joiner within Persian or Indic-script text, constitutes a different finding from the same character positioned between two English words.

3. Remove only unclassified residue

Remove the zero-width space, the word joiner, and any joiner or non-joiner not performing linguistic or emoji function. Retain the remainder.

No. These characters are present in text originating from PDF export, word processing software, translation tools, and ordinary web pages entirely unconnected to a language model. Identifying one establishes a fact about formatting or copying, not a fact about who or what produced the surrounding words.

A zero-width finding should be treated as a formatting fact warranting correction, not as evidence submitted toward a claim about a document's origin. These are distinct categories of claim, and only one is supportable by the character itself.

Zero-width characters have been used deliberately to divide a word within a hashtag without visually fragmenting it, to evade simple keyword filtering, and to cause two visually identical usernames to register as distinct strings under byte-level comparison. None of these applications concerns AI-generated text, and awareness of them prevents overweighting a single flagged character.

Where the text contains emoji, confirm compound sequences continue to render as single glyphs rather than as separated constituent images. Where the text is in Persian or an Indic script, confirm the letterforms display correctly. A removal pass that exchanges an invisible defect for a visible one has not achieved its objective, regardless of the resulting character count.