Invisible characters from ChatGPT and Claude: find them and strip them

Which characters are involved

• U+200B zero-width space, U+200C and U+200D joiners: invisible between two letters, but present in the bytes.
• U+FEFF byte order mark, used as an invisible space mid-text.
• U+2060 to U+2064: invisible operators with no rendering at all.
• Variation selectors U+FE00–U+FE0F and U+E0100 onwards: designed to pick a glyph form, repurposed as a data channel.
• Tag characters U+E0000–U+E007F: the stealthiest channel, able to encode entire strings completely invisibly.
• Bidirectional controls (U+202A–U+202E, U+2066–U+2069) and space homoglyphs (thin space, narrow no-break space) standing in for a normal space.

How to tell whether your text contains them

By eye, you cannot — that is the whole point. Three practical hints: your editor's character count is higher than what you can see; a find-and-replace on a word fails even though the word is right there; the text wraps in a strange place.

The reliable method is analysis: a tool that walks the code points and flags every Cf-category character and every variation selector. A good report gives you the position, the code point and the class of each occurrence, rather than a vague "suspicious text".

The trap of brute-force cleaning

Deleting every invisible character breaks perfectly legitimate text. U+200D is required after a composed emoji: without it a family becomes four separate figures and a regional flag turns into two letters. U+200C separates letters in Persian, Arabic and Devanagari: removing it changes the spelling.

The right rule is contextual. Keep load-bearing invisibles: a joiner after an emoji base, a non-joiner inside a word in a script that needs it, the tag characters of a valid flag. Strip everything else. A "paranoid" mode can remove all of it, but it must be an explicit choice, never the default.

Clean, then verify

After cleaning, paste the clean text back into the analyser: if the report is empty, no carrier is left. That round-trip check is what separates a real removal from a marketing claim.

One important caveat: removing invisible characters does not make text undetectable as AI writing. That is a different, statistical family of marks living in the word choices themselves. The two problems are handled separately.