01
> source text
Drop in the text
Paste it or load a .txt, .docx or .pdf. Anonymous scans take up to 2,000 characters and the scan never leaves your browser.
TWO LAYERS, NOT ONE
Claude hidden characters vs the watermark, code point by code point
Zero width spaces and lookalike letters are real characters you can find and delete. The Claude watermark isn't a character at all. Here's how to tell them apart.
Scan my text free Updated
claude.ai
Today
Give me a tagline for a coffee brand.
Message Claude
MacBook Pro
9:41
ClaudeCleanLocal scanPASTED FROM CLAUDE
SCANNING
…invisible characters
U+200Bzero-width space1
U+202Fnarrow no-break space1
U+200Dzero-width joiner1
9:41
ClaudeCleanLocal scanPASTED FROM CLAUDE
SCANNING
…invisible characters
U+200Bzero-width space1
U+202Fnarrow no-break space1
U+200Dzero-width joiner1
Short answer
They're two separate layers. Anthropic says its watermark adds nothing to the text and uses no hidden characters; it lives in which words Claude picks. Hidden characters like U+200B or U+202F are real code points you can find with a scan and remove with find and replace, and removing them leaves the watermark where it was.
THE CHARACTER LIST
What hidden characters can show up in AI text?
These are the invisible and lookalike code points a scan looks for. None of them is the Claude watermark.
X-RAY A PARAGRAPH
Seven characters you cannot see in this text
The paragraph below holds real invisible and lookalike characters. Switch on x-ray, then tap one to inspect it.
As it looks on screen
Our cafe opens at 7 am. Order ahead in the app and skip the line. Pastries are baked fresh every morning, and the cold brew is made in smаll batches. Family packs 👨👩👧 feed four. Prices include tax and tip.
Copy the paragraph with x-ray off and the characters come with it. The emoji family is three emoji glued by two ZWJs.
Inspector
U+202F
NARROW NO-BREAK SPACE
NNBSP
UTF-8 bytes
e280af
ClaudeClean
Replaced with a space
| Code point | Unicode name | What it does | ClaudeClean |
|---|---|---|---|
| U+200B | ZERO WIDTH SPACEZWSP | Invisible word separation and line break control, with no width. | Deleted |
| U+200C | ZERO WIDTH NON-JOINERZWNJ | Stops two letters from joining. Needed in Persian and some Indic scripts. | Deleted, mind the script |
| U+200D | ZERO WIDTH JOINERZWJ | Joins characters. It glues emoji ZWJ sequences like family emoji together. | Deleted, can split emoji |
| U+200E, U+200F | LEFT-TO-RIGHT and RIGHT-TO-LEFT MARKLRM, RLM | Invisible direction hints for mixed Latin and Arabic or Hebrew text. | Deleted |
| U+2060 | WORD JOINERWJ | A zero width non-breaking space, the preferred way to say 'don't break here'. | Deleted |
| U+FEFF | ZERO WIDTH NO-BREAK SPACEBOM | A byte order mark at the start of a data stream. Mid-text, it's leftover junk. | Deleted |
| U+00AD | SOFT HYPHENSHY | An invisible marker for a possible hyphenation spot. | Deleted |
| U+00A0 | NO-BREAK SPACENBSP | Looks like a space, stops a line break between two words. | Replaced with a space |
| U+202F | NARROW NO-BREAK SPACENNBSP | A thin no-break space. Reported in some o3 and o4-mini answers in 2025. | Replaced with a space |
| U+2007 | FIGURE SPACE | A no-break space as wide as a digit, used to align numbers. | Replaced with a space |
| U+202A to U+202E, U+2066 to U+2069 | BIDI EMBEDDING, OVERRIDE AND ISOLATE CONTROLS | Change the order text is displayed in without changing the stored order. | Deleted |
| U+FE00 to U+FE0F | VARIATION SELECTORSVS1 to VS16 | Pick a glyph variant. VS16 asks for the emoji style of a symbol. | Deleted, can change emoji |
| U+0430, U+0435, U+043E, U+0440, U+0441 | CYRILLIC а, е, о, р, сHomoglyphs | Look identical to Latin a, e, o, p, c but are different letters. | Swapped to Latin inside Latin words |
U+200B
ZERO WIDTH SPACE ZWSP
Invisible word separation and line break control, with no width.
ClaudeClean: Deleted
U+200C
ZERO WIDTH NON-JOINER ZWNJ
Stops two letters from joining. Needed in Persian and some Indic scripts.
ClaudeClean: Deleted, mind the script
U+200D
ZERO WIDTH JOINER ZWJ
Joins characters. It glues emoji ZWJ sequences like family emoji together.
ClaudeClean: Deleted, can split emoji
U+200E, U+200F
LEFT-TO-RIGHT and RIGHT-TO-LEFT MARK LRM, RLM
Invisible direction hints for mixed Latin and Arabic or Hebrew text.
ClaudeClean: Deleted
U+2060
WORD JOINER WJ
A zero width non-breaking space, the preferred way to say 'don't break here'.
ClaudeClean: Deleted
U+FEFF
ZERO WIDTH NO-BREAK SPACE BOM
A byte order mark at the start of a data stream. Mid-text, it's leftover junk.
ClaudeClean: Deleted
U+00AD
SOFT HYPHEN SHY
An invisible marker for a possible hyphenation spot.
ClaudeClean: Deleted
U+00A0
NO-BREAK SPACE NBSP
Looks like a space, stops a line break between two words.
ClaudeClean: Replaced with a space
U+202F
NARROW NO-BREAK SPACE NNBSP
A thin no-break space. Reported in some o3 and o4-mini answers in 2025.
ClaudeClean: Replaced with a space
U+2007
FIGURE SPACE
A no-break space as wide as a digit, used to align numbers.
ClaudeClean: Replaced with a space
U+202A to U+202E, U+2066 to U+2069
BIDI EMBEDDING, OVERRIDE AND ISOLATE CONTROLS
Change the order text is displayed in without changing the stored order.
ClaudeClean: Deleted
U+FE00 to U+FE0F
VARIATION SELECTORS VS1 to VS16
Pick a glyph variant. VS16 asks for the emoji style of a symbol.
ClaudeClean: Deleted, can change emoji
U+0430, U+0435, U+043E, U+0440, U+0441
CYRILLIC а, е, о, р, с Homoglyphs
Look identical to Latin a, e, o, p, c but are different letters.
ClaudeClean: Swapped to Latin inside Latin words
Names and abbreviations from the Unicode code charts. The ClaudeClean column describes what the cleaner does by default; it also catches any other format or control character, while keeping tabs and line breaks.
The definitions come from the Unicode General Punctuation chart and the Latin-1 Supplement chart. The BOM entry follows the Unicode BOM FAQ, which says U+2060 WORD JOINER is strongly preferred over ZWNBSP for word joining.
Lookalike letters are tracked in Unicode's UTS #39 confusables data. Cyrillic а is U+0430, not Latin a at U+0061, even though most fonts draw them the same.
SIDE BY SIDE
Is the Claude watermark a hidden character?
No. Here is how the character layer and the statistical layer differ on every point that matters.
| Question | Hidden characters | Claude watermark |
|---|---|---|
| What is it? | Extra code points stored in the string | A pattern in which words Claude chose |
| Can you see it? | Not on screen, but code, hex views and scanners show it | No, and no tool you can run shows it |
| Who can detect it? | Anyone, with an editor or a scanner | Only Anthropic's key, via a private-preview API |
| Does copy and paste keep it? | Usually, though some apps normalize spaces | Yes, it travels with the words |
| How do you remove it? | Find and replace, or a cleaner | Only by changing the words; Anthropic says a full rewrite does it |
| Does it identify you? | No | No, Anthropic says it carries no identifying information |
Swipe to compare →
Claude watermark details from Anthropic's explainer. Hidden character behavior from the Unicode charts and the scanner rules described on this page.
SKIP THE REGEX
Every code point in that table, flagged in one paste.
ClaudeClean scans for zero width characters, odd spaces, bidi controls, variation selectors and Cyrillic or Greek lookalikes, then cleans them, all in your browser.
- 300 rewrite credits a month
- 20,000 characters per request
- .txt, .docx and .pdf files
THE O3 EPISODE
Do ChatGPT hidden characters work like a watermark?
This is where the confusion started. A 2025 discovery about spaces got called a watermark long before any lab shipped one.
In April 2025, the startup Rumi reported that answers from OpenAI's o3 and o4-mini contained U+202F narrow no-break spaces, mostly in long responses and not in GPT-4o, according to Trending Topics. OpenAI made no public announcement, but Rumi later wrote that OpenAI told it the characters were not an intentional watermark.
Those spaces were visible in editors like VS Code and Sublime, and a find and replace took them out. That's the defining trait of the character layer: anyone can see it and anyone can clean it.
OpenAI's real text watermark arrived on October 5, 2026. Per BleepingComputer, textGrain is a statistical signal in word choice with no hidden characters, invisible spaces or unusual punctuation.
Don't treat odd spaces as proof
A U+202F in a document is not evidence that an AI wrote it. French typesetting, word processors, PDFs and web pages all produce no-break spaces of their own.
Going the other way, a clean string proves nothing either. Both Anthropic and OpenAI now mark text through word choice, which a character scan can't see. For the side-by-side, read ChatGPT vs Claude watermark.
APRIL 2025, REENACTED
Visible to an editor, gone after one replace
9:41
ChatGPTo3
How should I train for my first 10K?
Build up over 8202Fweeks. Run 3 times a week, with one long run that grows by about 1202Fkm each week. Keep most runs easy enough to talk. In the last week, cut the distance in half and rest 2202Fdays before the race. On race day, start slower than feels right for the first 2202Fkm.
4 x U+202F narrow no-break space
Find and replace
Find\u202FReplacespace
Replace all (4)Message
Sample answer written for this page. The character is real: U+202F, as Rumi reported in o3 and o4-mini output.
DO IT BY HAND
How to find hidden characters in AI text yourself
Six steps from a quick editor check down to raw bytes. You only need the first two for most documents.
- 01
Open the text in VS Code as code
VS Code 1.63 and later highlight uncommon invisible characters in source code by default, but Markdown isn't highlighted by default. Paste into a new file and set the language mode to a code language like JavaScript. - 02
Search with a regular expression
Open Find, switch on regex mode and search for [\u00A0\u00AD\u2007\u200B-\u200F\u202A-\u202F\u2060\u2066-\u2069\uFE00-\uFE0F\uFEFF]. Every match is a character from the table above. - 03
Confirm it in a hex view
Run xxd or hexdump -C on the saved file. In UTF-8, ZWSP shows as e2 80 8b, NBSP as c2 a0, NNBSP as e2 80 af, a soft hyphen as c2 ad and a BOM as ef bb bf. - 04
List code points in a browser console
Paste the text into a variable and map each regex match to its code point with codePointAt(0).toString(16). You get a list like U+200B, U+00A0, U+202F. - 05
Decide replace or delete per character
Swap no-break spaces for normal spaces, delete zero width spaces and soft hyphens, and leave joiners alone inside emoji or non-Latin scripts. - 06
Or let a scanner do all five
ClaudeClean runs the same checks in your browser, flags lookalike Cyrillic and Greek letters too, and applies those replace or delete rules for you.
Browser console snippet
const re = /[\u00A0\u00AD\u2007\u200B-\u200F\u202A-\u202F\u2060\u2066-\u2069\uFE00-\uFE0F\uFEFF]/g; [...text.matchAll(re)].map((m) => "U+" + m[0].codePointAt(0).toString(16).toUpperCase().padStart(4, "0"));
Run on "Hello", ZWSP, "world", NBSP, "and", NNBSP, "more", it returns U+200B, U+00A0 and U+202F. It doesn't cover lookalike letters, which need a confusables map.
STEPS 3 AND 4, RUN FOR REAL
What the hidden characters look like in bytes
The example from the snippet below, saved to a file and dumped with xxd. Highlighted bytes are the invisible characters.
zsh, sample.txt
$ printf 'Hello\u200Bworld\u00A0and\u202Fmore' > sample.txt
$ xxd sample.txt
00000000: 4865 6c6c 6fe2 808b 776f 726c 64c2 a061 Hello...world..a
00000010: 6e64 e280 af6d 6f72 65 nd...more
$ node snippet.js # the console snippet below
[ 'U+200B', 'U+00A0', 'U+202F' ]
3 hidden characters in 25 bytes.
# lookalike letters need a confusables map
$
MacBook Pro
$ printf 'Hello\u200Bworld\u00A0and\u202Fmore' > sample.txt
$ xxd sample.txt
00000000: 4865 6c6c 6fe2 808b Hello...
00000008: 776f 726c 64c2 a061 world..a
00000010: 6e64 e280 af6d 6f72 nd...mor
00000018: 65 e
$ node snippet.js # the console snippet below
[ 'U+200B', 'U+00A0', 'U+202F' ]
3 hidden characters in 25 bytes.
# lookalike letters need a confusables map
$
U+200BZWSPe2808bZERO WIDTH SPACE
U+00A0NBSPc2a0NO-BREAK SPACE
U+202FNNBSPe280afNARROW NO-BREAK SPACE
U+00ADSHYc2adSOFT HYPHEN
U+FEFFBOMefbbbfZERO WIDTH NO-BREAK SPACE
THE FAST WAY
An invisible character remover that shows its work
Paste once and see each hidden code point before anything is deleted.
02
scan report
3 zero-width1 soft hyphencleanable signals2Read the code point report
Zero width characters, odd spaces, soft hyphens, bidi controls, variation selectors and lookalike letters each get counted.
03
before → after
clearusefulanswerfocusedhelpfulresponse
Copied ✓Clean, then rewrite if needed
Cleaning handles the character layer. For the word-choice layer, an optional rewrite runs through a non-Claude model with a diff.
PRICING
Clean the text free. Rewrite more with Pro.
The scan and hidden-character cleanup are free. Pro adds room for rewrites and files.
PRO
CLAUDECLEAN PRO
ClaudeClean Pro
$5/ month
Cancel anytime.
- 300 credits every month (≈300,000 characters of rewriting)
- 20,000 characters per request
- Clean .txt, .docx and .pdf files
- Light / medium / strong intensity
- Dashboard with history
Checking account…
Free: 10 credits a month with an account, or try 2,000 characters right here without one. Strong rewrites cost 2× credits.
BEFORE YOU DELETE
Which invisible characters are safe to remove?
In plain English prose, nearly all of them. A few carry real meaning in other contexts.
Safe to strip in prose
- ZWSP, WJ and soft hyphens. They only hint at line breaks and hyphenation.
- A BOM mid-text. The Unicode FAQ defines it as a marker at the beginning of a data stream.
- Bidi controls in English. Latin-only text has nothing for them to reorder.
- NBSP and NNBSP. Replace them with a normal space so words don't run together.
- Lookalikes inside Latin words. A Cyrillic о in an English word is almost never intended.
Think before stripping
- ZWJ in emoji. Unicode's ZWJ sequences build single emoji from several, and removing it splits them.
- VS16 after symbols. Without it, some symbols fall back to their plain text style.
- ZWNJ in Persian or Indic text. It controls whether letters join, which changes how words read.
- LRM and RLM in Arabic or Hebrew. They keep punctuation and numbers in the right spot.
- Real Cyrillic or Greek words. ClaudeClean leaves words with no Latin letters alone for this reason.
LOOKALIKE LETTERS, MAGNIFIED
Same shape, different letter
Pick a pair. The lens reads each letter's code point, and only the Cyrillic one lights up.
The word on screen
spаce U+0073 Latin
Inside a Latin word spaceSwapped to Latin a, U+0061
A real Russian wordдомLeft alone: no Latin letters
LatinaU+0061LATIN SMALL LETTER A61
CyrillicаU+0430CYRILLIC SMALL LETTER Ad0b0
Most fonts draw both the same. The bytes differ: one byte for the Latin letter, two for the Cyrillic one in UTF-8.
THE KEY POINT
Does removing hidden characters remove the Claude watermark?
No. Cleaning the character layer and changing the word-choice layer are separate jobs.
Anthropic's explainer says nothing is added to the text and there are no hidden characters. There's nothing in the character layer for a cleaner to find.
The watermark only changes when the words change. Anthropic says light editing probably won't remove it, while a full rewrite that replaces every word will.
A cleaner still earns its place. Stray NBSPs break search and word counts, lookalike letters break spell check, and bidi controls can scramble how text displays.
What ClaudeClean does and doesn't do
It removes or normalizes the characters in the table above, in your browser. Its optional rewrite changes wording through a non-Claude model and shows a diff.
It can't read any provider's watermark, and rewriting is best effort. No detector result is promised. For the editing side, see does paraphrasing remove the Claude watermark.
TWO SHEETS, TWO JOBS
A cleaner reaches the top sheet only
Hidden characters sit on top of the text. The watermark is the text. Try both buttons.
Top sheet: characters
Quarterly revenueNBSP grew steadilyZWSP across everySHY region.
Bottom sheet: word choice
Quarterlyrevenuegrewsteadilyacrosseveryregion.
Hidden characters3 foundAny scanner sees them. Find and replace removes them.
Claude word choicesStill thereAnthropic says a full rewrite that replaces every word removes the watermark. Light edits probably won't.
Sample sentence written for this page. Tinting marks Claude's word choices for illustration; nobody outside Anthropic can see the real signal.
ZWSP, UP CLOSE
What is a zero width space used for?
U+200B has real jobs outside AI text. Here's what it does, how to type one and how to find one.
PURPOSE
Invisible break points
Unicode describes ZWSP as intended for invisible word separation and line break control. It gives a long URL or compound word a place to wrap without showing a space.
TYPING ONE
Escapes, not keystrokes
Most keyboards have no key for it. Write
​ in HTML or \u200B in JavaScript, Python or JSON strings.FINDING ONE
Three bytes in UTF-8
A ZWSP is stored as e2 80 8b. Search for those bytes in a hex view, or for \u200B in a regex-enabled find box.
For one-click copy, typing shortcuts and removal steps in VS Code, Word and Python, see the zero width space guide. To list every hidden code point in a text, use the invisible character detector.
WHY YOU CAN'T SEE IT
Watch why a text watermark leaves no visible trace
An 18-minute explainer that starts with red and green word lists and ends at SynthID-Text.
REFERENCES
Where do these Unicode definitions come from?
Primary references for every code point and claim on this page.
Unicode
General Punctuation chart (U+2000 to U+206F)
ZWSP, ZWNJ, ZWJ, LRM, RLM, NNBSP, figure space, bidi controls and word joiner, with their annotations.
View sourceUnicode
Latin-1 Supplement chart (U+0080 to U+00FF)
No-break space, commonly abbreviated NBSP, and the soft hyphen, abbreviated SHY.
View sourceUnicode
UTF and BOM FAQ
What a byte order mark is and why U+2060 is preferred over U+FEFF for word joining.
View sourceUnicode
UTS #39: Security mechanisms
The confusables data behind lookalike letter detection, such as Cyrillic а versus Latin a.
View sourceVisual Studio Code
November 2021 release (1.63)
Default highlighting of invisible and confusable characters, and the settings that control it.
View sourceAnthropic
Claude text watermark explainer
Where Anthropic states that nothing is added to the text and there are no hidden characters.
View sourceFor what the watermark is and who can read it, start at the Claude watermark guide.
FAQ
Claude hidden characters FAQ.
Anthropic says its watermark adds nothing to the text and uses no hidden characters. If a Claude draft contains zero width or no-break spaces, treat them as a separate formatting issue and check where the text passed through.
In April 2025 a startup reported U+202F narrow no-break spaces in o3 and o4-mini output, mostly in long answers; OpenAI later told the startup they were not an intentional watermark. OpenAI's October 2026 text watermark says it uses no hidden characters.
No. The watermark lives in word choice, so deleting zero width spaces or swapping no-break spaces leaves it untouched. Only changing the words themselves affects it. No detector result is promised.
U+200B, which Unicode abbreviates ZWSP. It has no width and marks an invisible word separation or a spot where a line may break. In HTML you can write it as ​ and in JavaScript as \u200B.
Since version 1.63, VS Code highlights uncommon invisible characters in source code by default. Markdown isn't highlighted by default, and the editor.unicodeHighlight settings switch each check on or off.
Mostly, in English prose. Zero width joiners build emoji sequences, ZWNJ matters in Persian and some Indic scripts, and no-break spaces should become normal spaces instead of disappearing.