Invisible Character Detector & Remover
Paste text or open a file to reveal zero-width spaces, hidden Unicode tag text, direction controls, look-alike letters and unusual spaces. Each one is marked where it occurs, and you can copy a cleaned version or download a cleaned Word file.
Hover over or tap a marker to see the character's name and Unicode code point. Dashed markers are characters that are normal in their place, such as the joiners inside family emoji.
The check runs in this page and does not upload your text. For Word files, the open-source JSZip library is downloaded from cdnjs to read the .docx package.
See the Characters You Cannot See
Text can look perfectly normal and still contain characters that take up no space, look exactly like other characters or change how the text is displayed. They slip in when you copy from web pages, PDFs, Word, chat apps and messaging tools, and they cause problems that are hard to trace: a search that misses a word that is obviously there, a password or username that will not match, a JSON file that fails at position 0, a command that breaks because a hyphen is really an en dash. This checker lists every such character with its official Unicode name, code point and line and column, marks it in place and gives you a cleaned copy.
What the checker looks for
| Type | Examples | Why it matters |
|---|---|---|
| Zero-width characters | Zero-width space U+200B, word joiner U+2060, byte order mark U+FEFF | Split words invisibly, so searches, comparisons and parsers fail |
| Unicode tag characters | U+E0020–U+E007E | Invisible copies of ASCII letters that can spell out a hidden message; the checker decodes it |
| Bidirectional controls | Right-to-left override U+202E, isolates U+2066–U+2069 | Make text display in a different order from how it is stored (the “Trojan Source” technique in code) |
| Variation selectors | U+FE00–U+FE0F, U+E0100–U+E01EF | One after an emoji is normal; long runs can hide data, which the checker decodes when it is readable |
| Look-alike letters and symbols | Cyrillic “а” in an English word, the Greek question mark “;” | Used in phishing links and to slip text past filters |
| Non-standard spaces | No-break space U+00A0, narrow no-break space U+202F, thin and em spaces | Break number parsing, splitting on spaces and code |
| Control and broken characters | NULL, ESCAPE, unpaired surrogates, noncharacters | Corrupt CSV imports, databases and terminals |
| Other formatting characters | Soft hyphen U+00AD, line separator U+2028, replacement character � | Invisible hyphenation points and line breaks that some apps ignore |
| Typographic punctuation | Curly quotes, en and em dashes, ellipsis, fullwidth letters | Visible, but break commands, code and CSV files that expect plain ASCII |
Hidden instructions for AI tools
Researchers have shown that some AI models read text written with Unicode tag characters even though people cannot see it. That makes it possible to hide instructions in a résumé, an email, a web page or a document that an AI assistant later processes — a technique known as ASCII smuggling. When the checker finds tag characters it decodes them and shows the hidden text in a red box. It also decodes runs of variation selectors and of zero-width characters when they spell out readable text, and for Word files it lists text formatted as hidden, coloured white or set in a font of 2 pt or smaller.
Characters that belong where they are
Not every invisible character is a problem. The zero-width joiner glues family and profession emoji together, the zero-width non-joiner shapes Persian and many Indian scripts, right-to-left marks keep Arabic and Hebrew text in order, and tag characters spell out the flags of England, Scotland and Wales. The checker recognises these from the characters around them, shows them with a dashed outline and, while Keep characters that are needed is ticked, leaves them in the cleaned text.
Does AI-written text contain hidden watermarks?
Text copied from chat apps can contain the same no-break spaces, curly quotes and em dashes as text copied from any web page or word processor, and the checker lists them. Statistical watermarks, such as Google's SynthID Text, are carried by the choice of words rather than by hidden characters, so removing characters does not affect them. This tool cleans text; it does not detect whether a person or an AI wrote it, and a few unusual characters prove neither.
Checking Word documents
Open a .docx file to check the text of the body, headers, footers, footnotes, endnotes, comments, image descriptions (alt text) and document properties. Download cleaned Word file rewrites only the characters in that text, and in tracked deletions and field codes, according to your options; fonts, styles, tables, images and layout stay as they were. Text hidden by formatting is listed for you to review in Word rather than deleted. Older .doc files and PDFs are not supported.
Limitations
- Look-alike detection covers Latin letters mixed with Cyrillic, Greek or Armenian ones in the same word, plus a few symbols that imitate punctuation. Whole words written in another alphabet are not flagged, and Unicode has thousands of other confusable characters.
- Fancy “bold” or “script” letters from the Unicode math range are not flagged; NFKC normalization turns them back into ordinary letters.
- Up to 2,000,000 characters are checked at a time, and the marked-up view shows the first 200,000.