Skip to main content
AI Watermark Removal

Local checker

Claude watermark checker

Paste text to find and remove invisible Unicode characters, with a count for each one. Use it for Claude copy/paste cleanup, code formatting, document hygiene, and pre-publication review.

Read this first: this cleans invisible characters, which is a real and separate problem. It does not remove Anthropic's watermark, and no method you can verify does, because the only detector for text is in a private preview limited today to eligible organizations under EU law.

Invisible character checker

This local tool finds and cleans common invisible characters before publishing, editing, or review.

Analysis runs in your browser.

Characters checked

0 found

50 kinds of invisible character are checked. Anything found appears here.

How to read this result

No result yet

Paste text above. Nothing leaves your browser.

Local processing

Your text never leaves this page, and you can prove it

  1. You paste text

    It goes into a React state variable in your browser tab. No form submits, no request fires.

  2. The scan runs locally

    A regular expression over the string, in your tab's JavaScript. No API, no server, no worker.

  3. Results render in place

    The counts, the diff, and the cleaned output are all computed from the same in-memory string.

  4. Analytics sees a number, never the text

    One event records how many characters were found and a coarse length bucket. The text itself is never a parameter, and there is no endpoint that accepts it.

  5. Verify it yourself

    Open your browser devtools, switch to the Network tab, and paste again. No request carries your text, because none exists.

What this tool checks

The checker scans for 50 kinds of invisible Unicode character: zero-width spaces and joiners, byte-order marks, soft hyphens, invisible math operators, the tag characters used to hide a message inside an ordinary sentence, directional controls, and the exotic spaces from em space down to hair space. These cause formatting, linting, search indexing, and copy/paste problems, especially in code, docs, CMS editors, and shared drafts. Characters doing real typographic or scripting work are flagged rather than deleted, because removing those corrupts your text.

What it does not do, and how we know

This tool does not remove Anthropic's watermark. Most pages selling that promise assume the mark is hidden Unicode, so we tested the assumption instead of arguing about it: 96 outputs across three Claude model tiers, 19,364 words, every code point counted. Zero zero-width or bidirectional-format characters, verified against a control proving the capture path preserves them.

So there was nothing of that kind to strip. That also means stripping it could never have been the thing that defeats Claude's mark. The data, the control, and the script are published in full, so you can check the claim rather than take it.

What Anthropic says

Anthropic's support article says supported Claude models weave an imperceptible watermark directly into generated text, and that a detected mark only means content may have been processed by Claude. On 14 August 2026 Anthropic explained the mechanism: a version of the SynthID-Text approach, where a secret key and the words that come before pick among meaning-preserving candidate words. "Nothing is added to the text and there are no hidden characters." That post also promised a detection API, which shipped on 1 September 2026. Rechecked on 5 September 2026: the exact algorithm and key remain unpublished, and the API is in private preview for eligible organizations under EU law rather than the public, a limit Anthropic's 4 September 2026 notice to administrators summarises as eligible EU organizations. That notice also says Anthropic does not see or store text submitted to the detection API, which matters if you were ever going to paste a passage into it, and which Anthropic's public help page still does not say when re-read on 25 September 2026. So a newsroom or a regulator can now check a passage, and you cannot, which means no removal claim aimed at you can be checked either.

One more detail worth keeping straight. Anthropic marks models "launched on or after August 2, 2026," and for the first thirty days its own release notes dated every generally available Claude model earlier than that. Fable 5.1 and Mythos 5.1 shipped on 1 September 2026 and Anthropic names them as supported, so the rule finally has models behind it. Until 4 September 2026 the retrofit for earlier models had no named model and no date. An email to Claude for Work administrators, seen that day, named Claude Opus 5 and 9 September 2026. A 25 September 2026 email to Claude Platform customers gives 14 September 2026 for Opus 5 instead and schedules Fable 5, Sonnet 5 and Opus 4.8 from 30 September 2026. Anthropic's help page, re-read that day, lists Opus 5 as marked. The help page now says every model released before 2 August 2026 will be covered by 2 December 2026; the full sourcing is on the Claude tracker.

FAQ

What does this checker find and clean?

Invisible Unicode characters and formatting artifacts that travel with copied text. Byte-order marks, soft hyphens, invisible math operators, and the tag characters used to hide a message inside an ordinary-looking sentence are deleted. Zero-width spaces and joiners are deleted where they are decorative and kept where they are doing typographic work. Exotic spaces, from the em space to the hair space, are folded back into a normal one. That is useful hygiene for code, documents, CMS editors, and anything that gets linted or diffed. It is not the same thing as removing Anthropic's watermark.

Why does it flag some characters instead of removing them?

Because invisible does not mean unwanted, and deleting the wrong character corrupts your text. Directional controls are how Arabic and Hebrew mix with English, and they are also how the Trojan Source attack makes source code read differently from how it runs. Private-use characters are how icon fonts work. A zero-width joiner is junk between two Latin letters and load-bearing between two emoji or inside a Devanagari conjunct, so that one is decided by what sits either side of it. Where a character has a real use, the tool points at it, explains it, and leaves your text alone, because you know what your document is for and we do not.

Does it remove Claude's actual watermark?

No, and neither does anything else you can currently buy. On 14 August 2026 Anthropic described the mechanism: a statistical watermark in which a secret key steers the choice among candidate words, explicitly not hidden characters, so a character cleaner has nothing of it to act on. Anthropic's own text detector shipped on 1 September 2026 into a private preview for the organizations EU law obliges to verify: regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups, plus enterprises with their own verification duty. Its 4 September 2026 notice to administrators summarises that limit as "eligible EU organizations". Unless you belong to one of those you cannot verify whether a passage carries the mark, which means you cannot verify its removal either, and any tool claiming otherwise is claiming a result you cannot check.

Does it matter when the text was generated?

Yes, and more than it used to. Anthropic marks models launched on or after 2 August 2026, which for a month meant only Fable 5.1 and Mythos 5.1. Anthropic's help page, re-read on 25 September 2026, now also marks Opus 5.5 and Opus 5. Anthropic's 25 September 2026 email to Claude Platform customers says Opus 5 has carried the mark since 14 September 2026 (a 4 September notice to administrators had said 9 September) and that Fable 5, Sonnet 5 and Opus 4.8 are scheduled to carry it from 30 September 2026. The help page says every earlier model will be covered by 2 December 2026. So the date a passage was generated changes whether it could carry the mark at all, and this checker cannot tell you either way, because it reads characters and the mark is not made of characters.

Does Anthropic say Claude uses zero-width spaces as the watermark?

No, and since 14 August 2026 it says the opposite outright: "Nothing is added to the text and there are no hidden characters." We tested the theory before Anthropic ruled on it: across 96 outputs and 19,364 words from three Claude model tiers, zero zero-width or bidirectional-format characters appeared.

So why use this at all?

Because invisible characters are a real, separate problem. They break linters and diffs, corrupt search and indexing, produce mysterious failures in code, and survive copy-paste between tools. Cleaning them is worth doing on its own terms. It is not worth doing under the belief that it defeats a provider's watermark.

Is the text uploaded?

No. The checker runs entirely in your browser and sends nothing to a server. You can verify that yourself: open your browser's network panel while using it and you will see no request carrying the text.

Sources and citation status