Hidden Characters in ChatGPT Text: Find and Remove Them

What zero-width and narrow no-break spaces in AI text do to hashtags, links and character counts, how to find them, and which ones to keep.

A social post with its invisible characters shown as small highlighted boxes between the words, next to the same post cleaned

Guide · By PostBoost team · · Markdown

Short answer: hidden characters in ChatGPT text are Unicode characters that take no space or look like an ordinary space: mostly the zero-width space (U+200B) and the narrow no-break space (U+202F). They are not a watermark. They matter on social media because a zero-width space inside a hashtag ends the hashtag, one inside a link ends the link, and on X some of them count as 2 characters. Paste the text into the free AI text cleaner to remove them, then check the count in the social media character counter. Keep the zero-width joiners inside emoji, or family and other combined emoji fall apart.

Where the hidden characters come from

The best-documented case is from April 2025. The education technology company Rumi reported that OpenAI's o3 and o4-mini models put narrow no-break spaces (U+202F) into longer answers such as essays, and suggested they were a watermark. OpenAI told Rumi two days later that "the special characters are not a watermark" and called them "a quirk of large‑scale reinforcement learning". On 23 April Rumi wrote that it was no longer seeing them.

So a fresh ChatGPT answer today may be clean. Hidden characters still turn up in AI drafts for ordinary reasons: the prompt itself was pasted from a web page, a PDF or a document that carried them, and the model or your editor kept them. Text copied from websites can hold non-breaking spaces (HTML  ), and hyphenated PDFs leave soft hyphens.

On watermarks more generally: OpenAI has said it developed a text watermarking method that it "continue[s] to consider", but it has not released one (Engadget, August 2024). An invisible character would be a poor watermark anyway, since one find and replace removes it.

The characters to look for

Character Code What it is for Cleaner does
Zero-width space U+200B Marks a possible line break in scripts without spaces, such as Thai Removes
Word joiner U+2060 Stops a line break between two characters Removes
Byte order mark U+FEFF Marks the start of a file; invisible in the middle of text Removes
Soft hyphen U+00AD Shows a hyphen only if the word breaks at the line end Removes
Mongolian vowel separator U+180E Mongolian script Removes
Non-breaking space U+00A0 A space that keeps two words on one line Turns into a normal space
Narrow no-break space U+202F The same, narrower: 9 am, 10 000 Turns into a normal space
Thin, hair and other spaces U+2002–U+200A, U+205F Typesetting widths Turn into a normal space
Zero-width joiner U+200D Glues emoji into one: family, profession and skin-tone emoji Keeps
Zero-width non-joiner U+200C Needed to spell words in Persian and other scripts Keeps

The cleaner also leaves the Braille pattern blank (U+2800) alone. That is the spacer the Instagram line break generator puts on blank lines on purpose, so a caption you format there keeps its gaps if you clean it afterwards.

What they do to a social post

We ran one post with five hidden characters through PostBoost's free tools: two zero-width spaces (one after "live", one inside #springmenu), two narrow no-break spaces (in "12 new" and "9 am") and one byte order mark.

Spring menu is live: 12 new dishes from $9, served from 9 am. Book a table on our site. #springmenu #brunch

On screen the two versions look the same. In the counter they do not:

Network With hidden characters Cleaned
X 113 107
Instagram, LinkedIn, Threads, Bluesky 110 107
YouTube description (bytes) 120 107

Three effects are worth knowing:

  1. The hashtag breaks. X's open-source twitter-text rules allow letters, numbers, the underscore and a short list of joiners in a hashtag. The zero-width space is not on that list, so #spring + U+200B + menu is the hashtag #spring followed by the word "menu". PostBoost's counter reads it the same way: it still finds two hashtags, but the first one is #spring.
  2. The link ends early. We put a zero-width space inside https://postboost.co/tools/ai-text-cleaner. On X a link counts as 23 characters; with the hidden character it counted 37, because the link stops at the zero-width space and the rest is plain text. Readers get a link to the wrong page.
  3. The count goes up. Under twitter-text weights, characters outside a few Unicode ranges count as 2 on X. The zero-width space is inside the "1" range; the narrow no-break space, word joiner and byte order mark are not, so each costs 2. YouTube counts descriptions in bytes, and each of these characters is 3 bytes in UTF-8. Five hidden characters are not much, but an essay-length AI draft cut down to a post can carry dozens. See the X character limit for how the 280 is counted.

How to find hidden characters

The quick way: paste the text into the AI text cleaner. With Invisible characters ticked, the summary under the result says how many it changed. If it reports no invisible-character changes, the text had none of the characters it removes or replaces.

In code: list every format character and every space that is not a plain space, with its position and name. This Python works on any text:

import unicodedata

def hidden(text):
    return [(i, f"U+{ord(c):04X}", unicodedata.name(c, "?"))
            for i, c in enumerate(text)
            if c != " " and unicodedata.category(c) in ("Cf", "Zs")]

print(hidden("Spring menu is live\u200b: 12\u202fnew dishes #spring\u200bmenu 👨‍👩‍👧"))

It prints the two zero-width spaces and the narrow no-break space, and also two ZERO WIDTH JOINER entries at the end: those belong to the family emoji. That is why a script that deletes every format character is the wrong fix.

How to remove them without breaking emoji

  1. Open the free AI text cleaner and paste the text.
  2. Leave Invisible characters ticked. If you only want this fix, set Em dashes to keep dashes and untick Curly quotes to straight, Double and trailing spaces, … to ... and Markdown ** and _ markers.
  3. Click Copy clean text and paste it where you schedule the post.
  4. Check the length in the character counter, or run the whole post through the post checker.

The cleaner removes zero-width spaces, word joiners, byte order marks, soft hyphens and the Mongolian vowel separator, and turns non-breaking, narrow and thin spaces into normal spaces. It keeps the zero-width joiner and non-joiner, so emoji and Persian words stay intact. Find and replace in a doc can do one character at a time if you paste it into the Find box, but you cannot see what you pasted, which makes it easy to miss one.

Clean AI text automatically with the API or MCP

If a script or an AI agent writes your posts, clean the text before it reaches the scheduler. The same cleaner is a free endpoint with no token:

curl -s https://postboost.co/app/api/tools/ai-text-cleaner \
  -H "Content-Type: application/json" \
  -d '{"text": "Spring menu is live\u200b: 12\u202fnew dishes #spring\u200bmenu", "invisible": true, "dashes": "keep", "quotes": false, "markdown": false}'

The response has the cleaned text and a changes list; for our full example post it was [{"type":"invisible","count":5}]. On the free tools MCP server at https://postboost.co/mcp/tools it is the clean-text tool, so an agent can call it in the same conversation in which it writes the post. The API docs list every option. For the dashes AI text leaves behind, see how to remove em dashes from ChatGPT text; if emoji in your post turn into empty boxes, that is a different problem, covered in emoji showing as boxes.

Do this in PostBoost

Free tools for this guide

FAQ

Does ChatGPT add hidden characters to its text?

Sometimes it has. In April 2025 the education technology company Rumi found narrow no-break spaces (U+202F) in longer answers from OpenAI's o3 and o4-mini models. OpenAI told Rumi they were not a watermark but a quirk of large-scale reinforcement learning, and three days after its first report Rumi said it no longer saw them. Hidden characters also come in with text copied from web pages, PDFs and documents.

Are hidden characters a ChatGPT watermark?

No. OpenAI said the characters Rumi found were not a watermark. OpenAI has said it developed a text watermarking method that it continues to consider, but it has not released one. Removing invisible characters does not make a text human-written either: it only removes characters you cannot see.

Do invisible characters count toward the character limit?

Yes. In PostBoost's counter a zero-width space counts as 1 character on Instagram, LinkedIn, Threads and Bluesky. On X, a narrow no-break space, a word joiner or a byte order mark counts as 2. In a YouTube description, which is counted in bytes, each of those characters takes 3 bytes.

Why does my hashtag stop in the middle of the word?

There is probably a zero-width space inside it. X's open-source twitter-text rules allow letters, numbers, the underscore and a few joiners in a hashtag, not a zero-width space, so #spring followed by a zero-width space and menu becomes the hashtag #spring. Delete the character or retype the hashtag.

Should I remove the zero-width joiner too?

Not inside emoji. The zero-width joiner (U+200D) glues emoji together: a family emoji is three or four people joined by it. Remove it and the emoji falls apart into separate people. Languages such as Persian also need the zero-width non-joiner (U+200C) to spell words correctly.

Is there an API to remove invisible characters?

Yes. PostBoost's AI text cleaner is POST https://postboost.co/app/api/tools/ai-text-cleaner, free and with no token needed, and the clean-text tool on the free tools MCP server at https://postboost.co/mcp/tools. With invisible set to true it removes zero-width characters and soft hyphens, turns non-breaking and thin spaces into normal spaces, and returns how many it changed.

Sources

Related guides

Try PostBoost free

API, webhooks and MCP on every plan, including Free. No credit card required.