Invisible Unicode Characters: The Complete Reference List

What counts as an invisible character

An invisible Unicode character is a code point that renders with no visible glyph, or with a glyph so small or so similar to a regular space that it is functionally invisible to a reader. Some exist for legitimate typographic or linguistic reasons — joining letters in Arabic, marking text direction, spacing out digits. Others exist purely as technical artifacts of software, and a few are used deliberately to hide text from human readers while keeping it machine-readable. The table below covers the ones you are most likely to run into, what each is actually for, and whether it is safe to strip on sight.

Reference table

Character	Code point	Category	Where it comes from	Safe to remove?
Zero Width Space	U+200B	Formatting	AI chat UIs, PDFs, web copy	Yes
Zero Width Non-Joiner	U+200C	Script control	Persian, Arabic, Indic script rendering	No — required in those scripts
Zero Width Joiner	U+200D	Script control	Emoji sequences, Indic script rendering	No — keep inside emoji
Word Joiner	U+2060	Formatting	Rich-text editors, AI chat UIs	Yes
Byte Order Mark / ZWNBSP	U+FEFF	Encoding marker	File headers, some text exports	Only legitimate as the first character of a file
Left-to-Right Mark	U+200E	Bidi control	Mixed-direction text, some chat UIs	Usually yes, outside bidi text
Right-to-Left Mark	U+200F	Bidi control	Mixed-direction text, some chat UIs	Usually yes, outside bidi text
LTR/RTL Embedding & Override	U+202A–U+202E	Bidi control	Legitimate bidi text; also "Trojan Source" code attacks	Flag and inspect — security risk
Bidi Isolates	U+2066–U+2069	Bidi control	Modern bidi-aware text rendering	Usually yes, outside bidi text
Soft Hyphen	U+00AD	Typography	PDFs and web pages at line-wrap points	Yes
No-Break Space	U+00A0	Spacing	Web pages, word processors, AI chat UIs	Yes — convert to a regular space
Narrow No-Break Space	U+202F	Spacing	French typography, some locales	Case by case
Figure / Thin / Hair Space	U+2007, U+2009, U+200A	Spacing	Typeset documents, number alignment	Case by case
Ideographic Space	U+3000	Spacing	CJK text and full-width input	No — required spacing in CJK
Mongolian Vowel Separator	U+180E	Script control	Mongolian script text	No — required in that script
Invisible Separator / Times / Plus	U+2062–U+2064	Math markup	MathML and technical publishing exports	Yes, outside math content
Hangul Filler	U+3164	Script control	Korean input method placeholders	Case by case
Variation Selectors	U+FE0E / U+FE0F	Presentation control	Emoji vs. text presentation switching	No — required for emoji presentation
Tag Characters	U+E0000–U+E007F	Deprecated tagging	Flag emoji; also abused for "ASCII smuggling"	Flag and inspect — security risk
Private Use Area	U+E000–U+F8FF	Application-defined	Icon fonts, app-specific glyphs	Case by case

How they get into text

Most invisible characters arrive as a byproduct rather than by intent. AI chat interfaces (ChatGPT, Claude, Gemini) use zero-width characters internally for cursor tracking and streaming, and these leak into copied output. Rich-text editors like Word, Google Docs, and Notion embed formatting metadata as invisible characters. PDF-to-text extraction routinely inserts non-breaking spaces and soft hyphens at line-wrap boundaries. Copying from a web page pulls in whatever non-breaking spaces and joiners the page's CSS and markup relied on for layout. A smaller set — bidi overrides and tag characters — can also be inserted deliberately, either to obscure text from a casual reader or, in security contexts, to smuggle instructions past a filter that only checks visible content.

Why they matter

For search and text matching, an invisible character splits what looks like one word into two tokens, or makes two visually identical strings fail an equality check. For code, they produce syntax errors and reference errors that are nearly impossible to spot by reading the source. For data, they break deduplication, VLOOKUP-style matching, and database joins on values that display as identical. For security, bidi override characters (U+202A–U+202E) can reorder how a line of code displays without changing what it executes — this is the basis of "Trojan Source" attacks — and tag characters (U+E0000–U+E007F) can encode hidden text inside what looks like a single emoji, a technique used in prompt-injection attempts against AI systems.

How to detect them

The raw view on our tools labels every hidden character it finds with its code point, so you see exactly what is in a string instead of guessing. Outside of that, a regex character class targeting specific code points (for example /[\u200B\u200C\u200D\uFEFF]/g in JavaScript) finds them programmatically, and most modern code editors — VS Code in particular — have a Unicode highlighting setting that flags unusual characters as you type.

What GPT Cleanup removes, and what it leaves alone

To be precise about scope: GPT Cleanup removes U+200B, U+200C, U+200D (except inside emoji sequences, where it is kept), U+2060, U+FEFF, U+200E, U+200F, and U+00AD, and it converts U+00A0 to a regular space. It does not currently touch the bidi override and isolate characters (U+202A–U+202E, U+2066–U+2069), the narrow and specialty spaces (U+202F, U+2007–U+200A), the ideographic space (U+3000), the Hangul filler (U+3164), tag characters, or private-use characters. If you are specifically checking for the security-relevant categories — bidi overrides or tag characters — you need to look for them separately; this list is meant to make that easy to do by hand or with a regex.

For a hands-on comparison across the common set, our zero-width space remover strips the categories listed as removable above, and our AI detector scans and reports on the full set found in a given piece of text. For background on why AI chat interfaces specifically leave these characters behind, see our companion post on zero-width characters.