AI Text Cleaner: Remove Hidden Characters in Word
November 11, 2026 · 11 min read
AI Text Cleaner: Remove Hidden Characters and ChatGPT Formatting
An AI text cleaner removes the invisible and leftover characters that ride along when you copy from ChatGPT or another chatbot. Think zero width spaces, no-break spaces, soft hyphens, citation markers and stray Markdown symbols. It doesn't rewrite or humanize anything. Below you'll find each character, how to spot it in Word, and how to strip it yourself.
What Does an AI Text Cleaner Actually Remove?
Pasted AI output usually carries five kinds of junk: invisible format characters, unusual spaces, typographic punctuation, chat leftovers such as 【3†source】, and Markdown symbols like ** and ###. The Markdown shows up because chatbots answer in it (we explain why in what Markdown means in AI tools).
Here's what an AI text cleaner, sometimes sold as a ChatGPT formatting remover, usually targets. The Word Find code column is what you type into Word's Find and Replace box.
| Character | Code point | Word Find code | What goes wrong |
|---|---|---|---|
| Zero width space | U+200B | ^u8203 | Searches miss, strings that look equal don't match |
| Zero width non-joiner | U+200C | ^u8204 | Needed in Persian and some Indic text, so check first |
| Zero width joiner | U+200D | ^u8205 | Holds emoji sequences together, so keep it |
| Word joiner | U+2060 | ^u8288 | Blocks a line break you expected |
| Byte order mark | U+FEFF | ^u65279 | Corrupts the first CSV header or fails a JSON parse |
| No-break space | U+00A0 | ^s | Looks like a space, fails exact matches |
| Narrow no-break space | U+202F | ^u8239 | Same problem, slightly thinner |
| Soft hyphen | U+00AD | ^u173 | Breaks searches for the word it sits in |
| Em dash | U+2014 | ^+ | Fine in prose, can trip up plain-text fields |
| En dash | U+2013 | ^= | Same as the em dash |
| Curly double quotes | U+201C, U+201D | ^u8220, ^u8221 | Break code, JSON and shell commands |
| Curly single quotes | U+2018, U+2019 | ^u8216, ^u8217 | Same, and the apostrophe is U+2019 |
| Ellipsis | U+2026 | ^u8230 | One character instead of three dots |
Most names and code points come from the Unicode General Punctuation chart. The no-break space and soft hyphen sit in the Latin-1 Supplement block instead. Note the two joiners: the Unicode emoji standard builds multi-person and profession emoji with U+200D, so a cleaner that deletes every zero-width character can break them.
Is There a ChatGPT Watermark in Copied Text?
Two separate stories sit behind this question, and most pages mix them up. Neither one is something you can see or delete in a Word document.
The first is about hidden spaces. In April 2025, Rumi reported that OpenAI's o3 and o4-mini models were leaving narrow no-break spaces (U+202F) in their answers. Two days later Rumi added an update: OpenAI said the characters weren't a watermark but "a quirk of large-scale reinforcement learning". By 23 April 2025, Rumi's tests no longer found them.
The second story is a real watermark. On 5 October 2026, OpenAI announced textGrain, which it says slightly changes the model's word choices to leave a statistical pattern. According to OpenAI, it's rolling out to eligible ChatGPT and Codex text in the EU. API developers can opt in for supported models, and it's off by default. Access to the detector starts with approved researchers and expert organisations.
So how does the ChatGPT watermark work? Through word choice, not characters. There's nothing to find in Word, and no text cleaner touches it. OpenAI also says the absence of a detected watermark doesn't prove human authorship. Hidden characters are a cleanup problem, not a provenance system. Running an AI text cleaner over a draft doesn't change who wrote it.
Where Zero-Width and No-Break Spaces Cause Real Damage
Invisible characters matter most when text stops being prose and becomes data. A zero width space inside a product code means a database lookup fails, even though both strings look identical on screen. A byte order mark at the start of a CSV file can turn the first column header into a different key.
Other common breakages:
- Search and replace. An exact-match search or script looking for
10 kgwon't find10andkgjoined by U+00A0. - Slugs and form fields. A CMS or sign-up form may reject a value with a hidden character in it.
- Code you paste into an editor. A no-break space in Python source can stop the file with a syntax error.
- Word counts. A zero width space can split one word into two for some counters.
This short script lists what it finds, then deletes the format characters and swaps the two no-break spaces for ordinary ones. It deliberately leaves U+200C and U+200D alone.
import re, sys
from collections import Counter
DROP = "\u200b\u2060\ufeff\u00ad" # zero width space, word joiner, BOM, soft hyphen
SPACES = "\u00a0\u202f" # no-break space, narrow no-break space
text = sys.stdin.read()
for ch, n in Counter(c for c in text if c in DROP + SPACES).items():
print(f"U+{ord(ch):04X} x{n}", file=sys.stderr)
text = re.sub(f"[{DROP}]", "", text)
text = re.sub(f"[{SPACES}]", " ", text)
sys.stdout.write(text)
Run it as python3 clean_ai_text.py < draft.txt > clean.txt. We tested it on a line holding a zero width space, a narrow no-break space, a no-break space and a family emoji. It reported this and kept the emoji's joiners intact:
U+200B x1
U+202F x1
U+00A0 x1
How to See ChatGPT Hidden Characters in Word
If you typed how to see ChatGPT watermark in Word into a search box, what you can actually see is hidden characters. Word shows some of them once formatting marks are on, and it finds the rest by code.
- On the Home tab, click Show/Hide ¶, or press Ctrl+Shift+8 on Windows or Cmd+8 on a Mac.
- Look at the gaps between words. A normal space shows as a small dot, and a no-break space shows as a small raised circle, a bit like a degree sign.
- Zero-width characters are easy to miss even with marks on, so search for them. Open Find and Replace, type
^u8203in Find what and click Find Next. - On Windows, put the cursor straight after a character you're unsure of and press Alt+X. Word swaps the character for its hexadecimal Unicode code. Press Alt+X again to swap it back.
One detail catches people out. The number after ^u is decimal, not hex, so ^u8203 finds a zero width space and ^u200B doesn't.
How to Remove ChatGPT Formatting in Word
Removing ChatGPT formatting in Word takes two passes: one for the visible styling and one for the characters underneath it. People who search how to remove ChatGPT watermark in Word usually mean those characters, since textGrain lives in the wording and no Word setting reaches it.
Visible formatting first. When you paste, open Paste Options and choose Keep Text Only. That drops bold, fonts, headings and links. If the text is already in your document, select it and click Clear All Formatting on the Home tab (the letter A with an eraser). On Windows, Ctrl+Spacebar does a lighter version that clears manual character formatting only.
Characters second. Keep Text Only doesn't touch invisible characters, because they're text. Open Find and Replace (Ctrl+H on Windows, or Edit > Find > Advanced Find and Replace on a Mac) and run these one at a time:
Find what Replace with Removes
^s one space no-break spaces
^u8239 one space narrow no-break spaces
^u8203 nothing zero width spaces
^u65279 nothing byte order marks
^u173 nothing soft hyphens
^+ - em dashes (optional)
^= - en dashes (optional)
Keep Use wildcards turned off, because Word's ^u codes don't work in a wildcard search. If ^u173 finds nothing, try ^-, which is Word's own code for an optional hyphen. If the paste also brought unwanted line breaks, the same dialog can remove line breaks in Word.
Curly quotes need one setting changed first. In Word for Windows, turn off "Straight quotes" with "smart quotes" under File > Options > Proofing > AutoCorrect Options > AutoFormat As You Type. Otherwise Word can curl the quotes again as it replaces them. Then find " and replace it with the same straight ". Word matches both curly forms and writes back straight ones. Do the same with the single quote.
The watermark under Design > Watermark is a different feature. It's a page background like DRAFT, and Remove Watermark in that menu deletes it.
Em Dashes and Curly Quotes: Normalize Them or Keep Them?
Models produce em dashes and curly quotes because published writing uses them. They're correct typography, not a defect. The only question is where the text ends up.
We prefer keeping curly quotes and dashes in finished prose, and normalizing only for code, CSV, JSON, shell commands and plain-text form fields. Swapping dashes for hyphens doesn't make text human-written, and it often makes the prose harder to read.
When you do need plain ASCII punctuation, a translation table handles all of it in one pass:
PLAIN = str.maketrans({
"\u2014": "-", "\u2013": "-", # em dash, en dash
"\u201c": '"', "\u201d": '"', # curly double quotes
"\u2018": "'", "\u2019": "'", # curly single quotes
"\u2026": "...", # ellipsis
})
print("\u201cIt\u2019s done\u201d \u2014 pages 4\u20139\u2026".translate(PLAIN))
That prints "It's done" - pages 4-9..., with straight quotes, hyphens and three separate dots.
Citation Markers and Other Chat Leftovers
Some leftovers are visible but still don't belong in your document. Answers that cite files or web pages can carry markers like 【3†source】 or bracketed numbers like [1]. Links copied from ChatGPT search results can end in a utm_source=chatgpt.com tracking parameter. If you select an answer with the mouse instead of the copy button, a code block's language label and Copy code caption can come along too.
These four substitutions handle the first three:
import re
text = (
"Rates rose in 2024 【3†source】 and again in 2025 [1][2, 5].\n"
"Read https://example.com/report?utm_source=chatgpt.com today.\n"
"Or https://example.com/a?utm_source=chatgpt.com&page=2 instead.\n"
)
text = re.sub(r"\s*【[^】]*】", "", text) # 【3†source】 style markers
text = re.sub(r"\s*(\[\d+(?:,\s*\d+)*\])+", "", text) # [1] and [2, 5]
text = re.sub(r"([?&])utm_source=chatgpt\.com&", r"\1", text)
text = re.sub(r"[?&]utm_source=chatgpt\.com", "", text)
print(text)
The output keeps the sentences and the rest of each URL:
Rates rose in 2024 and again in 2025.
Read https://example.com/report today.
Or https://example.com/a?page=2 instead.
Be careful with the bracket rule in Markdown files. A reference-style link such as [docs][1] uses the same pattern, so check those before you run it.
Strip the Markdown Last
Once the characters are clean, the **, ## and list dashes are all that's left. Paste the text into our Markdown to text tool, which runs in the browser and, like every browser-based tool on the site, is unlimited. Here's the limitation: it works on Markdown syntax and doesn't remove invisible Unicode such as zero width or no-break spaces, so do the character pass first.
For scripted methods, see converting Markdown to plain text. To keep real headings and bold in Word or Gmail instead, read Markdown to rich text. If one asterisk should survive as a literal character, escape it with a backslash before you strip the rest. For more consistent answers next time, set the output format in a reusable prompt template.
A good AI text cleaner pass runs in a fixed order: invisible characters first, then punctuation if the destination needs it, then chat leftovers, then Markdown. Turn on Show/Hide in Word to see what you're dealing with, and keep the decimal ^u codes handy. If you'd rather keep the Markdown, paste it into the free Markdown editor to check how it renders, with no account needed.