Short answer: they are two completely different things, and only one of them can be deleted.
Claude's text watermark is statistical. It is baked into how Claude picks its words — nothing is inserted into your text. Anthropic's own documentation states it directly: "Nothing is added to the text and there are no hidden characters." There is nothing in the file to strip. A zero-width-character cleaner, including the one on this site, cannot touch it.
Hidden characters are a separate layer. Zero-width spaces, directional marks, soft hyphens and similar code points are real bytes that can genuinely sit in a file. They arrive from many places — web pages, PDFs, CMS templates, Word autocorrect, and yes, some AI chat interfaces in some years. Those can be removed, and the check takes ten seconds.
| Claude's statistical watermark | Hidden / invisible characters | |
|---|---|---|
| What it is | A pattern in the choice of ordinary words, set by a key plus the preceding context | Specific Unicode code points (U+200B, U+FEFF, U+00AD, U+202F…) |
| Where it lives | In the sequence of words you can read | Between the characters you can read |
| Can you see it? | No — but neither can a detector without Anthropic's key | No, unless you turn on a control-character view |
| Can a cleaner delete it? | No. Deleting characters does nothing to it. | Yes — it is a byte, you can drop it |
| Who can verify it | Anthropic's detection API (private preview) — regulators, law enforcement, media, fact-checkers, researchers, education, EU civil society | Anyone, with the scripts below |
| Introduced | Models released on or after 2 August 2026; older models being added over the coming months | Decades old. Long predates AI |
If you landed here because a site promised to "remove the Claude watermark", read the next section before you trust it with your document.
Anthropic published How Claude's text watermarking works on 14 August 2026 (updated 1 September 2026). These are its own words, not a journalist's summary:
The mechanism is worth understanding, because it explains why deletion is impossible. A language model picks the next word from a list of candidates. Often several candidates are equally good — "overcast" or "grey" both finish the sentence, and it genuinely does not matter which. Normally that tie is settled by a random number generator. With watermarking, the tie is settled by a key plus a few preceding words instead.
So the words are still random. What changed is where the randomness comes from. A detector holding the key can replay the sequence and ask: "is this text consistent with the choices Claude would have made using this key?" That yields a probability, not a proof.
The consequence nobody selling a remover wants to state: because the signal lives in which readable words were chosen, there is no character to delete, no tag to strip, no metadata to wipe. Every "Claude watermark remover" that works by cleaning invisible characters is removing something that was never the watermark.
Three limits from the same document are widely ignored, and they matter if you are worried about being accused of using AI:
Anthropic is also explicit that translations carry the watermark, because every word is chosen by the model, and that light editing probably will not remove it while a complete rewrite where every word is replaced will.
They are real, they are common, and they have nothing to do with the 2026 watermark. The usual sources:
No upload, no account. This prints the offset, code point and Unicode name of every character in the Cf, Cc, Zl and Zp general categories — the four categories invisible formatting characters live in.
python -c "
import unicodedata, sys
text = open(sys.argv[1], encoding='utf-8').read()
hits = [(i, c) for i, c in enumerate(text)
if unicodedata.category(c) in ('Cf', 'Cc', 'Zl', 'Zp')]
print(len(hits), 'invisible character(s)')
for i, c in hits[:40]:
print('offset', i, 'U+%04X' % ord(c), unicodedata.name(c, 'UNKNOWN'))
" yourfile.txt
Save it as scan.py and run python scan.py draft.md. A clean file prints 0 invisible character(s). That is the entire "detection" step — there is no server involved and no AI model needed, because this layer is just bytes.
This is the part that gets lost. Anthropic's document also says that when Claude produces a file of a supported type — .png, .jpg, .svg — it attaches a C2PA content credential: a cryptographically signed note in the file's metadata saying the file was made or processed with Claude. Anthropic is explicit that this "is very different from a watermark. Nothing in the file changes — it is not embedded or hidden."
It is metadata. Metadata can be read, and metadata can be removed. Check for it with ExifTool:
exiftool -a -G1 -s image.png | grep -iE "c2pa|jumbf|provenance|claim_generator"
If that returns a JUMBF or C2PA block, the file carries a provenance credential. Stripping it is a normal metadata operation and the tool on this site does it in the browser, without uploading the image. Worth knowing what you lose: C2PA is the same standard cameras and photo editors use to record origin, so removing it also removes the record that you shot or edited the file. The byte-level container map, and the profile you should not delete →
Then be clear about which thing you want, because the honest answers are very different.
Before you bulk-delete: zero-width non-joiner (U+200C) and zero-width joiner (U+200D) are required characters in Arabic, Persian, Hindi, and several other writing systems — they change how words render and join. Directional marks (U+200E, U+200F, U+061C) are required in bidirectional text. A blanket strip will corrupt those files. Scan first, look at the code points you actually found, then decide.
Three reasons, and only one of them is dishonest.
That last point is the whole reason this page exists. The character layer is real, it is worth cleaning, and the C2PA layer is genuinely removable. But if a site will not tell you plainly that it cannot touch the statistical watermark, it is either misinformed about what it is deleting or hoping you will not check.
Anthropic's answer is no — "there are no hidden characters". If you find invisible characters in something you pasted from Claude, treat them as coming from your editor, the clipboard, or whatever the text passed through, not from the watermark. Run the scan and see which code points they are.
Not by reading the watermark. The key is Anthropic's, and detection needs it. AI detectors work by a completely different method — they look at phrasing tells and statistical properties of the text, which Anthropic itself notes is "fundamentally different from checking for a watermark". So a detector's verdict is not a watermark check, and removing characters will not change what a detector says.
No. The words are unchanged. The watermark is in the words, and you did not touch them.
Anthropic says the EU law includes a transition period for models launched before 2 August 2026 and that they are working to add watermarking to those as well, rolled out over the coming months. Watermarking is applied globally rather than by region because they do not yet have a durable way to scope it geographically.
Anthropic's document describes no opt-out, and no user-level setting. It also states the watermark carries nothing that identifies you, your organisation, or your chats.
Related: how to find the hidden characters in your own file · cleaning a whole folder, and blocking them in git · the browser cleaner