Strip hidden markers
from AI text.

Paste output from ChatGPT, Claude, Gemini or any other model. The scanner reports every invisible character, look-alike letter and typographic fingerprint it finds, then hands back clean plain text. Nothing leaves your browser.

01 PASTE TEXT

0 CHARS 0 WORDS

02 CLEANING PROFILE

AWAITING INPUT

0 FINDINGS
0 CHANGES
0 CHARS

REPORT

Paste text on the left, then press Convert.

  • HIDDEN
    Zero-width characters U+200B–U+200D, U+2060, U+FEFF. The most common carrier for a hidden marker, and a routine copy-paste artefact.
  • HIDDEN
    Soft hyphens U+00AD. Invisible until a line breaks, then it surfaces as a stray hyphen.
  • HIDDEN
    Unicode tag characters U+E0000–U+E007F. An entire ASCII message can be smuggled in this block with nothing visible on screen.
  • HIDDEN
    Variation selectors U+FE00–U+FE0F, U+E0100–U+E01EF. Legitimate after an emoji; arbitrary data when attached to a plain letter.
  • HIDDEN
    Blank filler glyphs Braille blank and Hangul fillers, characters that render as nothing but count as letters.
  • HIDDEN
    Line and paragraph separators U+2028 and U+2029. Real line breaks that almost nothing downstream handles, so they collapse a document into one line or split it in the wrong place.
  • HIDDEN
    Control characters C0 controls, DEL and the C1 block, excluding tab, newline and carriage return. They render as nothing, survive a copy, and corrupt CSV, JSON and database fields.
  • HIDDEN
    Invisible format characters U+2061-U+2064 invisible mathematical operators, and the deprecated U+206A-U+206F format block. No glyph, no width, and they survive a copy exactly as a zero-width space does.
  • HIDDEN
    Bidirectional controls U+202A–U+202E, U+2066–U+2069. The trojan-source trick: the text you read is not the text you copied.
  • SUBSTITUTED
    Non-breaking spaces U+00A0. Survives copy-paste and quietly breaks search, diffs and code.
  • SUBSTITUTED
    Narrow no-break spaces U+202F. Frequently seen in Gemini output, and also correct French typography.
  • SUBSTITUTED
    Typographic spaces En, em, thin and hair spaces (U+2000–U+200A, U+205F). A space to the eye, not to a computer.
  • HIDDEN
    Mixed-script look-alikes Cyrillic а е о р с, Greek ο ν and full-width forms standing in for Latin letters inside a Latin word.
  • STYLE
    Em dashes U+2014. The single most recognisable stylistic tell in GPT and Claude prose.
  • STYLE
    En dashes and minus signs U+2013, U+2011, U+2212 and relatives, all flattened to a plain hyphen.
  • STYLE
    Curly quotes and apostrophes Typographic quotes applied automatically by every major model.
  • STYLE
    Ellipsis character U+2026 standing in for three periods.
  • STYLE
    Bullet glyphs Literal bullet characters where a plain list marker was meant.
  • STYLE
    Repeated spaces Two or more spaces inside a line, collapsed to one. Indentation at the start of a line is preserved.
  • STYLE
    Trailing whitespace Spaces and tabs left at the end of a line.
  • STYLE
    Repeated blank lines Three or more consecutive newlines, collapsed to a single blank line.

MARKED

Highlighted spans are detected markers. Characters with no glyph are shown as their codepoint.

  
 
 
CLEAN TEXT
 
 
 
 
 
  
 HOW IT WORKS // Everything runs in this page. Your text is never uploaded, stored or sent anywhere. Findings are reported for the whole input; only the groups you leave enabled are actually changed. 
 
 
  
 WHAT IT CANNOT DO // Character-level cleaning does not defeat statistical watermarks, Google SynthID, or the token-level mark Anthropic has applied to Claude output since August 2026, because those are carried in word choice rather than in the characters themselves. Read the output before publishing, some of these characters are legitimate typography. 
 
 
 
    
 
 
 
 
  
 
 
 
 
TEXT TOOL
 
Questions about hidden text markers.
 
Character-level cleaning does not defeat statistical watermarks, Google SynthID, or the token-level mark Anthropic has applied to Claude output since August 2026, because those are carried in word choice rather than in the characters themselves. Read the output before publishing, some of these characters are legitimate typography.
 
  
 
 
  What is an AI text watermark?   
Two different things share the name. One is a character-level marker: an invisible codepoint, a look-alike letter, or a typographic habit that survives copy-paste and can be found by reading the bytes. The other is a statistical watermark encoded in which words a model chooses. This tool finds and removes the first kind.
 
  Does this remove Google SynthID or make text undetectable?   
No. SynthID, and the mark Anthropic has applied to Claude output since August 2026 under the EU AI Act, are carried in word choice rather than in the characters, so no character-level tool can strip them. This tool also makes no claim about the score any AI detector will give your text.
 
  Is my text uploaded anywhere?   
No. Detection and cleaning both run in your browser. The page makes no network request with your text, nothing is stored, and the tool keeps working with the network disconnected.
 
  Which models does it work on?   
All of them, because it reads the text rather than its origin. It has no way to know which model produced a string and does not claim to. What it reports is what the characters are.
 
  Why does it flag em dashes and curly quotes?   
They are legitimate typography, not hidden markers, but they are also the most recognisable habit of GPT and Claude prose. They sit in their own group so you can see them reported without having them removed.
 
  Will it damage text written in Russian, Greek or Chinese?   
No. A look-alike letter is only replaced inside a word that is otherwise Latin. A word written entirely in Cyrillic, Greek or any other script is left exactly as it is.