Dev.to AI ๐Ÿค– Ai ๐Ÿ‘ 0 ๐Ÿ“– 4 min read

What a "ChatGPT watermark remover" actually removes, and what it did to 100 AI scores

We built a browser-based ChatGPT watermark remover that strips hidden Unicode characters. Then we tested whether cleaning changed the scores from our AI detector. Short version: for 96 of 100 texts the score stayed with

We built a browser-based ChatGPT watermark remover that strips hidden Unicode characters. Then we tested whether cleaning changed the scores from our AI detector.

Short version: for 96 of 100 texts the score stayed within 5 points.

Two different things get called a "watermark"

A real text watermark is statistical. The model nudges which words it picks, and a verifier with the key can spot the pattern later. There are no extra characters to find. As of November 11, 2026:

  • OpenAI says verification of its text marks is available only to approved organizations, such as research and academic institutions.
  • Google watermarks text from the Gemini app and website with SynthID.
  • Anthropic says recent Claude models and some older ones add a text watermark.

No character cleaner can touch any of these, because there is nothing in the string to delete.

Hidden characters are a separate thing. Zero-width spaces, odd-width spaces and soft hyphens come along when you copy from a chat window, a PDF or a web page. They aren't a signature. They do break code, mess up line wrapping in a CMS and throw off word counts.

What our cleaner looks for

Character Code What it does
Zero-width space, non-joiner, joiner U+200B to U+200D Removed (the joiner is kept inside emoji like ๐Ÿ‘จโ€๐Ÿ‘ฉโ€๐Ÿ‘ง)
Word joiner, byte order mark U+2060, U+FEFF Removed
Soft hyphen U+00AD Removed
Thin, hair, en, em and other fixed-width spaces U+2000 to U+200A Replaced with a normal space
Narrow no-break space U+202F Replaced with a normal space
Cyrillic or Greek letters inside English words e.g. U+0430 ะฐ Replaced with the Latin letter
No-break space U+00A0 Kept, because it often belongs there, as in "Mr. Smith"

Find them yourself in 20 lines

You don't need our tool to see what's in a string. This is a cut-down version of the check, plain JavaScript, no dependencies:

const NAMES = {
  0x200b: "zero-width space", 0x200c: "zero-width non-joiner", 0x200d: "zero-width joiner",
  0x2060: "word joiner", 0xfeff: "byte order mark", 0x00ad: "soft hyphen",
  0x202f: "narrow no-break space", 0x205f: "medium mathematical space", 0x3000: "ideographic space",
};
const isOddSpace = (cp) => cp >= 0x2000 && cp <= 0x200a;
const isLookAlike = (ch) => /[\u0400-\u04FF\u0370-\u03FF]/.test(ch); // Cyrillic or Greek

function findHidden(text) {
  const hits = [];
  [...text].forEach((ch, i) => {
    const cp = ch.codePointAt(0);
    const name = NAMES[cp] ?? (isOddSpace(cp) ? "fixed-width space"
      : isLookAlike(ch) ? `look-alike letter "${ch}"` : null);
    if (name) hits.push({ index: i, code: "U+" + cp.toString(16).toUpperCase().padStart(4, "0"), name });
  });
  return hits;
}

const sample = "The res\u200Bults were cle\u00ADar:\u202Fthe m\u043Edel did better\u2009than expected.";
console.table(findHidden(sample));

Running it with Node prints:

โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”
โ”‚ (index) โ”‚ index โ”‚ code     โ”‚ name                    โ”‚
โ”œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ผโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ผโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ผโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ค
โ”‚ 0       โ”‚ 7     โ”‚ 'U+200B' โ”‚ 'zero-width space'      โ”‚
โ”‚ 1       โ”‚ 21    โ”‚ 'U+00AD' โ”‚ 'soft hyphen'           โ”‚
โ”‚ 2       โ”‚ 25    โ”‚ 'U+202F' โ”‚ 'narrow no-break space' โ”‚
โ”‚ 3       โ”‚ 31    โ”‚ 'U+043E' โ”‚ 'look-alike letter "ะพ"' โ”‚
โ”‚ 4       โ”‚ 46    โ”‚ 'U+2009' โ”‚ 'fixed-width space'     โ”‚
โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ดโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ดโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ดโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜

Printed, that sample looks ordinary. It contains a zero-width space, a soft hyphen, two unusual spaces and a Cyrillic "ะพ".

The look-alike check here is crude: it flags every Cyrillic or Greek letter, which is wrong for text that is actually Russian or Greek. The version on our page only flags them inside otherwise Latin words.

Does cleaning change an AI score?

We wanted to know whether it does, so we measured it on our own detector.

We took 100 AI-written texts and made a copy of each with hidden characters added: zero-width spaces, unusual spaces, soft hyphens and a few look-alike letters, about 59 per text. We scored the copy with hidden characters and the cleaned version 3 times each, without using saved results, and compared the averages.

Score change after cleaning (cleaned minus with hidden characters) Points out of 100
Median change 0
Middle 80% of texts -0.7 to 0
Largest drop and rise -33.0 and +0.3
Texts within 5 points 96 of 100
Texts whose verdict changed 3 of 100

Our detector only, engine version w0p5-38c6d591, November 11, 2026. Other detectors may behave differently; we didn't test them.

The other four texts dropped by more than 5 points; the largest drop was 33 points, and three verdicts changed. No text went up by more than 0.3. We averaged three runs per version because scores vary between runs. We didn't compare cleaning with changes to the wording.

When cleaning is worth it

  • Code. A zero-width space inside a variable name, or a thin space in an indent, gives you a syntax error you can stare at for an hour.
  • CMS and email. Pasted text that won't wrap, or wraps in the wrong place, often has fixed-width or no-break spaces in it.
  • Word and character limits. Invisible characters still count toward form limits and some word counters.
  • Search and matching. "results" with a zero-width space in the middle won't match "results".

If any of that sounds familiar, paste the text into a checker or run the snippet above before you go hunting for a bug that isn't in your logic.

I used AI to help draft this post.

๐Ÿ“ฐ Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes โ€” full credit and traffic to the original publisher.