Invisible characters in AI text

• Zero-width characters

  Zero-width space U+200B, zero-width non-joiner U+200C and joiner U+200D, word joiner U+2060, zero-width no-break space U+FEFF, Mongolian vowel separator U+180E, combining grapheme joiner U+034F.

  glyphwash: Removed. Joiners stay where they do a job: inside emoji, and between letters in scripts such as Arabic and Devanagari, where they change how letters connect.

• Unusual spaces

  No-break space U+00A0, narrow no-break space U+202F, thin and hair spaces U+2009 U+200A, figure and punctuation spaces U+2007 U+2008, en, em and smaller fixed-width spaces U+2002 to U+2006, medium mathematical space U+205F.

  glyphwash: Replaced with an ordinary space.

• Soft hyphens

  U+00AD, a hyphen that only appears when a word breaks at the end of a line.

  glyphwash: Removed.

• Text-direction controls

  Direction marks U+200E U+200F U+061C, embeddings and overrides U+202A to U+202E, isolates U+2066 to U+2069.

  glyphwash: Removed. In text with Hebrew or Arabic the marks, embeddings and isolates stay, since they're needed there. Overrides, which can make text display in a different order from how it's stored, never stay.

• Invisible maths operators

  Function application U+2061, invisible times U+2062, invisible separator U+2063, invisible plus U+2064.

  glyphwash: Removed.

• Tag characters

  U+E0000 to U+E007F, invisible copies of letters, digits and symbols that can spell out a whole message.

  glyphwash: Removed, and any message they spell out is shown to you. Emoji flags that use them, such as Scotland's, stay.

• Variation selectors

  U+FE00 to U+FE0F and U+E0100 to U+E01EF, which pick a style for the character before them.

  glyphwash: Removed, except one that styles an emoji or a Chinese or Japanese character. A run of them is how data hides in text, so a run always goes.

• Look-alike letters

  Letters from other alphabets posing as Latin ones, such as the Greek omicron U+03BF in "repοrt".

  glyphwash: Replaced with the letter they imitate.

• Copying and pasting. Web pages, PDFs and word processors are full of no-break spaces and soft hyphens, and they come along with the text.
• Model habits. In April 2025 some answers from OpenAI's o3 and o4-mini contained narrow no-break spaces, which OpenAI said were a quirk of training, not a watermark.
• Content credentials. The C2PA standard can embed a signed record of where a text came from inside the text itself, written as a zero-width no-break space followed by a run of variation selectors.
• Fingerprints and hidden messages. Zero-width and tag characters can encode an ID or a message that travels with the text wherever it's pasted, including instructions aimed at AI tools that read it.

• They can identify a text: the same invisible pattern shows where a copy came from.
• They break things: searches miss words with a zero-width space inside, code stops compiling, word counts drift.
• They can carry instructions: a message in tag characters is invisible to you but readable by an AI model.
• Look-alike letters make words that look right but don't match in searches or spell checks.

The watermarks Claude, ChatGPT and Gemini use aren't characters at all. They're patterns in which words were chosen, so a text with every hidden character removed still carries one. glyphwash takes the characters out first, then rewrites every passage with an open-weight model that adds no watermark of its own. How the Claude watermark works.

• Paste the text into glyphwash: the result lists each hidden character, where it was and what replaced it, and shows any hidden message.
• Open it in a code editor such as VS Code, which highlights invisible and look-alike characters.
• In code, this JavaScript pattern finds the invisible ones, though not unusual spaces or look-alike letters:

/[\u00AD\u034F\u061C\u180E\u200B-\u200F\u202A-\u202E\u2060-\u2064\u2066-\u2069\uFEFF\u{E0000}-\u{E007F}]/gu