Invisible Character Detector

Every character in a string is a numbered Unicode code point. The detector walks the text one code point at a time and checks each number against a list of about 60 characters that render as nothing or as blank space: zero-width spaces and joiners, the byte order mark, soft hyphens, Hangul fillers, the braille blank, no-break and odd-width spaces, direction controls, variation selectors, tag characters and stray control codes.

Normal spaces, tabs and line breaks are left alone. Everything else on the list gets a tag in the revealed view and a row in the table, with its U+ code point and official Unicode name so you can look it up.

Ask JavaScript for "hello".length and you get 5. Plant a zero-width space inside and you get 6, though the screen still shows five letters. That extra slot is why two identical-looking passwords fail to match, why a spreadsheet lookup misses, and why a character counter disagrees with your eyes.

It gets stranger. .length counts UTF-16 units, so an emoji or a tag character counts as 2. This tool counts real code points instead, and reports UTF-8 bytes separately, because that is what databases and SMS limits usually measure.

What you see as one character can be several code points glued together: a family emoji is four people joined by zero-width joiners, and a red heart is a plain heart plus variation selector 16. Those invisible parts are legitimate, so the detector skips a joiner that sits between two emoji and a variation selector right after one. Flag emoji for England, Scotland and Wales use tag characters and are skipped too.

Accents work the same way: “e” plus a combining acute is one letter. Real scripts rarely put more than two marks on a letter, so only runs of three or more are flagged as stacked marks, which is the Zalgo effect.

Cyrillic “а” and Latin “a” are different code points that look identical in most fonts. Swap one into a web address and you have a phishing link that passes a glance. The checker knows about 45 Cyrillic and Greek letters that pass for Latin ones.

It only flags them inside a word that also contains Latin letters, or in words made entirely of look-alikes when the text is mostly Latin. That keeps real Russian or Greek text from lighting up. It is a heuristic, not the full Unicode confusables table, so it can miss rarer swaps.

• Word and Google Docs. Soft hyphens, no-break spaces and zero-width joiners ride along when you copy from a document or a PDF.
• Web pages and CMSs. Byte order marks at the start of a pasted file, NBSPs from   in HTML, zero-width spaces inserted to allow line breaks in long URLs.
• AI chat output. Narrow no-break spaces and other odd spaces have turned up in chatbot answers. See the AI section below.
• Source code. Direction overrides and look-alike letters planted to make code read one way and run another.
• Game and social names. Braille blanks and Hangul fillers typed on purpose to get a blank or spaced-out username.

In 2025 people noticed narrow no-break spaces and other unusual spaces in answers from some ChatGPT models and called them a watermark. OpenAI said they were a quirk, not a tracking mark. Either way, they are easy to spot here and easy to remove.

Real AI watermarks, such as Google's SynthID for text, are statistical patterns in word choice. Stripping characters does nothing to those. If you want AI text tidied up beyond hidden characters, the AI text cleaner handles em dashes, markdown leftovers and the rest.

In 2021 Cambridge researchers showed that bidirectional controls like RIGHT-TO-LEFT OVERRIDE (U+202E) can make source code display in a different order from how the compiler reads it (CVE-2021-42574). A comment can appear to close when it does not, or a check can appear to run when it is skipped. The same trick turns invoice[RLO]fdp.exeinto a file name that reads “invoiceexe.pdf”. The example text above includes one.

Zero-width characters can encode bits. Give each recipient of a memo, a tweet draft or a leaked-looking post a unique pattern of them and any copy-pasted leak points back to its source. Tag characters go further and can hide a whole sentence, which is how instructions get smuggled into prompts for AI models. If you are about to repost text you did not write, run it through here first.

The friendlier use. Name fields usually trim spaces and zero-width characters, so people reach for characters that count as letters but draw nothing: the braille blank (U+2800) and the Hangul filler (U+3164). That is what the copy card above is for. The same characters are why a username can look empty in a chat log, and why this detector counts them as hidden.

Press Ctrl+Shift+8 (Cmd+8 on Mac) or click the ¶ button to show formatting marks. That reveals no-break spaces as a small degree sign and soft hyphens as a hooked dash, but not zero-width characters.

For those, open Find and Replace (Ctrl+H), type ^u8203 in Find what (the decimal code of U+200B) and leave Replace with empty. Repeat with 8204, 8205, 8288 and 65279. Use ^s for no-break spaces and ^- for soft hyphens.

Docs has no view that shows invisible characters. Open Edit → Find and replace, tick Match using regular expressions, search for [\u200B-\u200D\u2060\uFEFF] and replace with nothing. If the count says zero but something still feels off, paste the text into the detector above; it also catches look-alike letters, which no find box will.

VS Code highlights invisible and ambiguous characters by default since version 1.63, through the editor.unicodeHighlight.invisibleCharacters and editor.unicodeHighlight.ambiguousCharacters settings. Turn on editor.renderControlCharacters to see control codes too.

To remove them across a project, open search with regex on (Alt+R) and run the same character class as above, replacing with nothing. GitHub also shows a warning banner on files that contain bidirectional controls.

Cleaning up text that came out of a chatbot? The AI text cleaner strips hidden characters plus the stylistic tells. Tightening your own writing? The passive voice checker flags every sentence that hides who did what.