Client-side utility · v1
Strips the hidden characters and provenance metadata that mark text and images as AI-generated. Free, instant, and it runs entirely in your browser — nothing is uploaded.
New · 11 Aug 2026Anthropic announcement
Claude now watermarks everything it writes, with no opt-out. Anthropic says metadata goes when a file’s format changes — so half of this is genuinely solvable. Read which half →
0characters removed
Zero-width spaces, bidi controls, soft hyphens and Unicode tag characters deleted from pasted text.
Counted when someone copies the cleaned result, not on every keystroke.
C2PA, EXIF and XMP destroyed by re-encoding the pixels in the browser.
01Capability
AI content is marked in more than one way and no single technique clears all of them. Most tools publish only the first half of this table.
| Mark | Applies to | Removed |
|---|---|---|
| Invisible Unicode charactersZero-width spaces, bidi controls, soft hyphens, Unicode tag block | Text | YES |
| Exotic space charactersNarrow, hair, ideographic and non-breaking spaces | Text | YES |
| C2PA provenance manifestSigned manifest, destroyed by re-encoding the pixels | Images | YES |
| EXIF and XMP metadataCamera, GPS location and editing history | Images | YES |
| Statistical text watermarkLives in word choice, not characters — only rewriting affects it | Text | NO |
| Pixel-level image watermarkSurvives re-encoding, matched by perceptual hash | Images | NO |
02Mechanism
Conflating these is the most common mistake in this category. One is a container you can strip. The other is not a thing you can delete.
Attached to the file
An AI system can bolt a cryptographically signed record onto an image saying it was generated or edited by AI. C2PA is the open standard for it, and EXIF and XMP carry camera and editing history alongside. Because the signature covers the file's bytes, re-encoding the picture invalidates the whole thing.
Removed here
Baked into the wording
When a model writes, it picks each word from a spread of plausible options. A watermark nudges those picks along a hidden pattern. The prose reads normally and means the same thing, but across a long enough passage the bias is measurable. It is a property of the word sequence, which is why copying carries it and paraphrasing erodes it.
Not removed here
Deleting invisible characters does not change a single word, so it has no effect on a statistical watermark. Any tool claiming otherwise is describing the first column while selling the second.
Deep rewrite
Paraphrases text to break statistical watermarks. The only approach that touches the second column above — and it needs a server, so it cannot run on your device the way this tool does.
04Sponsors
Every slot is visible on every page view. Price rises with position: the early ones are cheap because the audience is unproven, the last is dearest because it is the last.
01$200/mo
Available →
02$300/mo
Available →
03$400/mo
Available →
04$500/mo
Available →
05$650/mo
Available →
06$800/mo
Available →
07$1000/mo
Available →
08$1150/mo
Available →
Pick a slot above, or write to [email protected].
05Reference
Including the ones with inconvenient answers. If a question below has a No in it, that is the real answer rather than a hedge.
30 of 30
01
Two things. From text, it deletes invisible Unicode characters — zero-width spaces, bidirectional controls, soft hyphens, Hangul fillers, and the Unicode tag block used to smuggle hidden messages — and normalises exotic spaces to ordinary spaces. From images, it strips signed provenance metadata: C2PA manifests, EXIF, and XMP. Both run entirely in your browser.
02
No. A statistical watermark is not a hidden character you can delete. It lives in which words the model chose — the model biases its token selection so the resulting sequence carries a signal a detector can measure. Deleting invisible characters does not change a single word, so it has no effect on that signal. The only thing known to degrade a statistical watermark is rewriting the text substantially, which is a different operation from cleaning it.
03
When a language model writes, it picks each word from a range of plausible options. A watermarking system nudges those choices according to a hidden pattern. The text still reads normally and means the same thing, but across a long enough passage the pattern is statistically detectable. It is a property of the word sequence itself, which is why copying and pasting carries it along and why paraphrasing weakens it.
04
C2PA is an open standard for content provenance. An AI system can attach a cryptographically signed manifest to an image recording that it was generated or processed by AI. Because the signature covers the file's contents, re-encoding the image destroys it. This tool redraws the image through a canvas and re-encodes it, so the manifest, along with EXIF and XMP, does not survive.
05
The text and images you clean are never uploaded. All of that processing happens in your browser using JavaScript and the canvas API, there is no account, and nothing you paste or drop is stored or transmitted. Three things do leave the page, none of them your content: Google Analytics and Cloudflare Web Analytics record ordinary page-view information such as your country and referring page, and the running totals shown on this site are kept up to date by sending three integers — how many texts were cleaned, how many images, and how many characters were removed. A Content-Security-Policy restricts outbound connections to those endpoints and nothing else.
06
No, and any tool claiming otherwise is overstating what it can verify. Removing a provenance mark removes one signal. It does not remove statistical watermarks in text, pixel-level watermarks in images, or stylistic patterns that classifiers look for. Equally, the absence of a mark was never proof that content is human-written — plenty of AI content carries no mark at all.
07
No. Some provenance systems pair signed metadata with a watermark embedded in the pixels themselves, which survives re-encoding and can be matched against a perceptual-hash database. Re-encoding removes the metadata container but leaves the pixel data essentially intact, so any mark carried in the pixels remains.
08
PNG, JPG, and WebP, up to 10 MB. The cleaned file is exported in the same format it came in, so a JPG stays a JPG. Anything else is rejected before processing.
09
It is free with no account and no limits. Because everything runs on your device there is no server cost per use, so there is nothing to meter.
10
Use the Inspect tab before cleaning. It lists every hidden character found in your text by code point, gives its Unicode name, counts it, and states what the cleaner will do with it. If Unicode tag characters are present it also decodes the smuggled message so you can read what was hidden. Nothing is modified in that view, so it works as an independent check rather than a claim.
11
The zero-width joiner is load-bearing inside emoji. Removing every occurrence turns a family emoji into four separate people and breaks skin-tone and profession variants. The cleaner deletes a joiner sitting between ordinary letters, where it can only be hiding something, and preserves one sitting between two pictographs, where it is doing real work.
12
Yes, and there is nothing Claude-specific about it. Invisible characters and exotic spaces come from the same sources regardless of which model produced the text, and the image side reads C2PA, EXIF and XMP the same way whatever wrote them. The same limits apply too: statistical watermarks from any provider are untouched.
13
Yes, after the first visit. The page is a static site with no server calls needed to clean anything, so once it is in your browser cache it works without a connection. The running totals will simply not update.
14
Text is processed in a single pass with no size limit beyond your browser's memory; five megabytes of text cleans in roughly thirty milliseconds. Images are capped at 10 MB, checked before any processing begins so an oversized file is rejected immediately rather than after a wait.
15
Re-encoding a JPEG or WebP is lossy, so a small amount of detail is lost. At the quality setting used here it is not visible in normal viewing, but it is real and it accumulates if you process the same file repeatedly. PNG is lossless and loses nothing. Output format always matches input, so a JPG stays a JPG.
16
U+200B, a character that occupies no horizontal space but is still a character. It was designed to mark where a long unbroken string may wrap across lines. Because it is invisible but real, it survives copy and paste, breaks exact-match search, and makes two visually identical strings compare as different.
17
Chat interfaces apply typographic touches when rendering a response, and one common choice is a narrow no-break space (U+202F) either side of an em dash to stop it wrapping awkwardly. The model never produced it; the rendering layer did. Copying from the rendered page picks it up. This tool normalises it to an ordinary space rather than deleting it, because deleting it would join the words together.
18
Not directly. It handles plain text and PNG, JPG and WebP images. For a document, copy the text out and paste it into the cleaner, then paste the result back. Note that Office and PDF files carry their own metadata — author, company, revision history — which this tool does not touch and which the application's own document-inspector feature can strip.
19
No. This tool handles text and still images only. Video watermark removal is a fundamentally different problem requiring frame-by-frame processing, and visible logo removal in particular is unrelated to the invisible marks discussed here.
20
Yes. The interface is built for touch, the image dropzone is tappable rather than drag-only, and processing happens on the device the same way it does on a desktop. Very large images may be slower on older phones because canvas re-encoding is memory-bound.
21
Not currently. The page does register WebMCP tools, so an AI agent running in a browser that supports the standard can call the cleaner directly. That standard is still in origin trial, so support is limited.
22
Not by this site directly. Cloudflare, which serves the pages, keeps standard edge request logs as any host does, and Google Analytics processes IP addresses to derive approximate location. Neither ever receives the text or images you clean, because that content never leaves your browser.
23
A humanizer rewrites your text to change how it reads, which alters your words and can introduce errors. This tool changes no words at all — it only deletes characters that were never visible and normalises spaces. If your text was already correct, it stays correct. The trade-off is that it cannot touch anything living in word choice, including statistical watermarks.
24
It should not, and it often fixes things. Invisible characters pasted into source code cause syntax errors that are almost impossible to spot by eye, and this removes them. Ordinary spaces, tabs and newlines are left untouched, so indentation and Markdown structure survive intact.
25
Google DeepMind's watermarking technology, applied across text, images, audio and video. For images it embeds a pattern in the pixels themselves rather than in metadata, which is why it survives cropping, compression and re-encoding. This tool removes metadata containers and cannot affect pixel-embedded marks of that kind.
26
The Act places transparency obligations on providers and deployers of AI systems, particularly for synthetic media and interactions where people should know they are dealing with AI. Whether a given piece of your content falls under those obligations depends on context and jurisdiction. This is general information, not legal advice — if it matters commercially, ask a lawyer.
27
There is no published policy singling them out, and they are common enough in ordinary web content that treating them as a ranking signal would be unreliable. The practical reasons to remove them are technical rather than SEO: they break search matching, corrupt data parsing, and produce confusing diffs.
28
Two ways. It publishes what it cannot do — the removal table lists pixel-level and statistical marks as not removed, which most tools omit. And the Inspect tab lets you verify the result yourself by listing every hidden character by code point, so the claim is checkable rather than something you take on trust.
29
Because a full account is more useful than a list of wins. A zero-width joiner between two emoji is doing legitimate work and is deliberately preserved, so the inspector shows it and marks it kept. If it only listed what gets deleted you could not tell the difference between a character that was left alone on purpose and one that was missed.
30
This tool is intended for personal privacy and formatting cleanup on content you own — stripping location data from a photo before sharing it, or removing invisible characters that break formatting. You remain responsible for complying with the EU AI Act, platform terms of service, and any academic or professional rules that require disclosing AI assistance.