ChatGPT Watermark Detector & Hidden Character Checker

ChatGPT watermark detector: what can actually be found in a text

People search for a ChatGPT watermark detector expecting a hidden signature that proves a text came from an AI. Here is what is real. OpenAI has said it built a text watermarking method and decided not to ship it, partly because rewording, translation or a second model wipes it out. Google does watermark Gemini output with SynthID Text, but that mark lives in the choice of words, and only Google's own tools can check it. Neither of these can be read by an independent checker, including this one.

What a checker can find is everything that sits in the text as characters: zero-width spaces, invisible format characters, look-alike letters from another alphabet and unusual spacing. These come from AI chat interfaces, from copying web pages and PDFs, and sometimes from people who tag copies on purpose. Paste the text, and the report lists each type it found, with a detail line under the badges, for example Found 2 zero-width/invisible characters.

Does ChatGPT add hidden characters?

Sometimes, but not as a declared watermark. In April 2025 users noticed narrow no-break spaces (U+202F) in answers from newer ChatGPT models and called them a watermark. OpenAI did not confirm that, and the character appears in ordinary typography too (French punctuation, numbers like 10 000). Zero-width spaces and soft hyphens in a pasted answer can also come from the app or page you copied it from, so the character alone does not tell you its origin.

Two practical conclusions:

• Finding an invisible character does not prove the text is AI-written. Not finding one does not prove the opposite.
• If you care about U+202F specifically, check it with the command below: this detector treats U+202F as a normal space and does not flag it.

For a judgement about authorship, use a style-based tool such as the AI text detector, and treat its score as a hint, not a verdict.

How to read the report

The badges are the finding types. Each one means a specific rule fired:

The confidence percentage is not a probability that the text is AI-generated. It is a formula that grows with the number of finding types and characters found: one zero-width character gives about 46%, several types quickly reach 100%. Read the detail lines, not the number.

Zero width space detector: the characters, one by one

Check text for hidden characters without uploading it

If the text is confidential, check it locally.

Microsoft Word

Show/Hide ¶ (Ctrl+Shift+8) reveals spaces, no-break spaces and optional hyphens, but not zero-width spaces. To find those, open Find and type ^u8203 (the decimal code of U+200B); ^u8204, ^u8205 and ^u65279 find ZWNJ, ZWJ and BOM. Replace with nothing to remove them.

Google Docs

Docs has no mode that shows zero-width characters. Paste the text into this checker or into a code editor, clean it there and paste it back.

VS Code and other code editors

VS Code highlights invisible Unicode characters by default (setting editor.unicodeHighlight.invisibleCharacters), so hidden characters show up as boxes.

Command line

This prints every line that contains a zero-width character, BOM, soft hyphen or U+202F, and works on macOS and Linux:

perl -CSD -ne 'print "$.: $_" if /[\x{200B}-\x{200F}\x{2060}-\x{2064}\x{FEFF}\x{00AD}\x{202F}]/' text.txt

And this lists each suspicious character with its position and Unicode name:

python3 -c "import sys,unicodedata as u
t=open(sys.argv[1],encoding='utf-8').read()
[print(i,hex(ord(c)),u.name(c,'?')) for i,c in enumerate(t) if u.category(c)=='Cf' or c in '  ­']" text.txt

Where hidden characters actually cause damage

• URLs and slugs. A zero-width space pasted into a link or a CMS slug produces an address that looks right and returns 404.
• Codes and logins. Promo codes, API keys and usernames copied from a chat fail validation with no visible reason.
• Source files. A BOM before <?php sends output before headers and triggers "headers already sent"; some JSON and CSV parsers choke on it or read it into the first column name.
• Search and counts. A word with a hidden character or a Cyrillic letter inside is a different word for Ctrl+F, site search and spell checkers, and word counters disagree.
• Titles and snippets. A soft hyphen in a title or meta description can surface as a visible hyphen when the text is reflowed elsewhere.

Removing hidden characters safely

The clean copy does more than delete zero-width characters. Know what changes before you publish it:

• Zero-width and format characters are removed, including the joiners inside emoji sequences and the non-joiners that Persian or Hindi spelling needs.
• Look-alike letters are replaced with the expected alphabet. A mixed word such as a Latin brand name with a Cyrillic suffix will be changed too.
• Runs of spaces collapse to one, and trailing spaces are trimmed.
• Strings caught as metadata_markers are deleted, including the phrases "AI-generated" and "Generative AI", UUIDs and long base64 strings. If your text discusses AI or contains IDs, restore them by hand.
• Em and en dashes become hyphens, curly quotes become straight quotes and no-break spaces become ordinary spaces. This happens silently and is not counted as a finding.

If you only want to strip invisible characters and keep typography, use the Word replace method or the Perl command with s///g instead.

What a clean result does not prove

"No watermarks detected" means none of the rules above fired. It does not mean the text is human-written, and it says nothing about statistical watermarks such as SynthID Text. Retyping, translating or paraphrasing a text removes character-level marks completely, which is exactly why providers moved to statistical ones.

Images, video and audio carry provenance differently, in file metadata such as C2PA manifests. For those, use the image watermark detector, the video watermark detector and the audio watermark detector.