The three answers that matter most:
1. Claude does not add hidden characters. Anthropic's own documentation: "Nothing is added to the text and there are no hidden characters." The 2026 watermark is a statistical pattern in how the words were chosen.
2. No character cleaner can remove it — including the tool on this site. There is nothing in the file to delete.
3. What can be removed is a different layer: invisible characters that really are bytes in your file, and the C2PA / EXIF / XMP metadata Claude attaches to images and files.
Eighteen questions below, each answered in a few sentences. Where a longer explanation exists, it is linked rather than repeated.
No. Anthropic's documentation states it plainly: "Nothing is added to the text and there are no hidden characters." The 2026 watermark is a statistical pattern in how the words were chosen, not characters inserted into the string.
If you do find invisible characters in something you pasted, they came from somewhere else — your editor, the clipboard, a PDF you copied from, or a web page. Which code points they are is worth checking, because the fix depends on it. How to list them in your own file →
Not by reading the watermark. Detection requires Anthropic's key. Their detection API is in private preview, available to organisations required to verify watermarking under EU law — regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups — plus enterprises with their own compliance obligations.
AI detectors work by an entirely different method: they look at phrasing tells and statistical properties of the text. Anthropic calls this "fundamentally different from checking for a watermark." So a detector's verdict is not a watermark check, and deleting characters will not change what it says.
No. The words are unchanged, and the watermark lives in the words. You removed a different layer entirely. The two layers explained →
No — and no editor setting will reveal it. It is a pattern in the choice of ordinary words, so there is nothing to see. Turning on "show formatting marks" in Word, or a control-character view in VS Code, will not expose it, because there is no character there to expose. Those settings are still useful, but for the other layer.
Anthropic says light editing probably will not remove it completely, while a complete rewrite in which every word is replaced will. In that second case they note it is arguable whether the text can still be described as AI-generated.
Their reasoning is mechanical, not policy: the watermark rides on ties between equally good word choices. Rewrite the words and you replace the ties. Edit a comma and you have not.
Being added. 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.
They also note watermarking is applied globally rather than by region, because they do not yet have a durable way to scope it geographically.
No opt-out is described in Anthropic's documentation, and there is no user-level setting. The same document states the watermark carries no identifying information and cannot be traced to a specific person, organisation, or chat — so it is not a tracking mechanism, which is the concern most people actually have when they ask this.
Yes. Anthropic says a translation produced by Claude carries a watermark, because in that case every word is chosen by the model.
Proofreading is the opposite case. When Claude only fixes grammar and punctuation, nearly all the words remain the author's, so there is very little for the watermark to attach to — and depending on length, Claude's involvement may not be detectable at all.
C2PA is signed metadata, and yes, you can remove it. When Claude produces a supported file type — PNG, JPG, SVG — it attaches a C2PA content credential: a cryptographically signed note in the file's metadata recording that the file was made or processed with Claude. Anthropic describes it as "very different from a watermark" and notes "nothing in the file changes — it is not embedded or hidden."
Because it is metadata rather than a statistical pattern, it is readable and removable. Check for it with ExifTool:
exiftool -a -G1 -s image.png | grep -iE "c2pa|jumbf|provenance"
A JUMBF or C2PA block in the output means the file carries a provenance credential. Worth knowing the trade-off: C2PA is the same standard cameras and photo editors use to record origin, so stripping it also removes the record that you produced the file. The full container map, and what else sits beside it →
Not in the way most articles describe. OpenAI's own provenance documentation describes no production text watermark; their provenance work covers images and audio, using SynthID and C2PA.
The "hidden characters" stories around ChatGPT, which circulated heavily in 2025, concern character artifacts — narrow no-break spaces and similar — rather than a provider watermark. Treat them as an interface artifact. Do not use them as a comparison point for Claude's 2026 watermark; the two are different kinds of thing. Cleaning those characters across a folder →
Because invisible characters landed in the file. The usual culprits are zero-width space (U+200B), zero-width joiner (U+200D), the byte-order mark (U+FEFF), and the line separators U+2028 / U+2029. They break diff, grep, CSV and JSON parsers, and some linters.
This is the character layer, not the watermark — and unlike the watermark, it is genuinely fixable. Scan the folder, then block it at commit time so it cannot come back. The folder scan and git hook →
Mechanically, you cannot — there is nothing in the file to remove. What you can do depends on which problem you actually have:
The browser cleaner handles the first two, locally, with no upload. It does not touch the third, and no tool does.
Four things, all of them real bytes: invisible and zero-width characters; bidirectional control marks; hidden HTML class attributes; and image metadata (EXIF, XMP, C2PA).
It cannot detect or remove a statistical watermark, and it does not rewrite or "humanise" your wording. If a site tells you otherwise about its own tool, ask it which layer it is operating on. The full comparison →
Depending on which layer you mean. If you mean the statistical watermark, no tool removes it — it is not stored in your text. If you mean invisible characters that came along with a copy-paste, those are real bytes and they do come out.
The honest split: Anthropic's own page says "Nothing is added to the text and there are no hidden characters." So anything promising to strip "the Claude watermark" out of a text file is either removing characters that were never the watermark, or rewriting your wording. Which layer is which →
Not publicly, as of now. Detection needs Anthropic's key, and access is described as a private preview for regulators, law enforcement, news organisations, fact-checkers, researchers and educational institutions. There is no public detector you can paste text into.
What you can check yourself is a different thing: whether the file carries invisible or zero-width characters. Those are visible to a scanner even though they are invisible on screen. How to look →
On how it works, yes: Anthropic describes a statistical approach in the SynthID-Text family. What changes is the source of the randomness used when the model picks the next token — a key plus the preceding text decides it. Nothing is inserted into the output.
On removal, the honest answer follows from that: deleting characters cannot touch it, because there are no watermark characters to delete. Rewriting the text completely produces a different piece of text, which is the only thing Anthropic's description leaves open. The mechanism, in plain terms →
Every time the model writes the next word, it picks from a shortlist. Normally it picks using randomness. With watermarking on, that randomness is decided by a secret key plus the words already written — so the pattern of choices carries a shape only the key can see.
The words themselves are ordinary words. That is why nothing shows up when you inspect the characters, and why no cleaner can find it by looking for hidden marks.
The watermark is not a ranking signal. Google's public position is that it judges content on quality rather than on how it was produced, so a page is not penalised for being AI-assisted — it is judged on whether it is useful.
Where the watermark does matter is outside search: the detectors a client, a school or an employer runs on a submission. That is a contract and policy question, not an SEO one.
One caution before you bulk-delete: zero-width non-joiner (U+200C) and zero-width joiner (U+200D) are required in Arabic, Persian, Hindi and several other writing systems, and directional marks are required in bidirectional text. A blanket strip will corrupt those files. Scan first, read the code points you found, then decide.
Sources: Anthropic, How Claude's text watermarking works (published 14 August 2026, updated 1 September 2026); Google DeepMind's SynthID-Text paper in Nature (2024). Anthropic's statements are quoted directly; anything not in their document is marked as our reading, not theirs.