Frequently asked questions
Everything people usually ask before trusting a cleaner with their text. For the underlying detail, see the invisible Unicode guide.
Does the tool upload my text anywhere?
No. Cleaning runs entirely in your browser with JavaScript. Nothing you paste or upload is sent to a server, logged or stored, and closing the tab discards it.
Can it remove SynthID or other statistical watermarks?
No. SynthID-class watermarks are embedded in a model's token choices, not in characters, so character-level cleaning cannot affect them. Only substantially rewriting the text perturbs that kind of signal.
Which characters does it detect?
Zero-width and invisible formatting characters, exotic and no-break spaces, bidirectional controls, Unicode tag characters (U+E0000–U+E007F), variation selectors, Cyrillic and Greek homoglyphs, optional stylized punctuation, and Markdown or HTML provenance artefacts such as attribution lines and utm_source tracking parameters.
Will cleaning change the meaning of my text?
The invisible categories are removed without any visible change. Space normalisation, homoglyph normalisation and typographic normalisation do alter characters, so each is reported with its exact codepoint and replacement, and the last two are optional toggles.
Is it safe for non-English text?
Turn off homoglyph normalisation for genuinely multilingual text — Cyrillic and Greek letters are legitimate there. Bidi controls are also legitimate in right-to-left documents. Always review the detection report before accepting the output.
Can I clean a file instead of pasting?
Yes. Upload or drag a .txt file onto the input area. You can download both the cleaned text and a plain-text audit report listing every codepoint that was changed.
Can it remove watermarks from PNG images?
Not yet — binary metadata in PNG files is coming soon. Today the cleaner operates on text only, so it does not read image files or touch C2PA content credentials, EXIF tags or steganographic signals embedded in pixel data. When PNG support ships it will strip provenance metadata (C2PA manifests, EXIF software/source fields) and re-encode the image to drop hidden byte-level markers, without altering visible pixels. It will not attempt to defeat pixel-level statistical watermarking.
Can it strip provenance from PDF or DOCX files?
Not yet — PDF and DOCX metadata cleaning is on the roadmap. Those formats embed author, creator, generator and revision identifiers in document properties and XMP packets, plus AI-assistant attribution lines in body text. The text inside a PDF or DOCX can be extracted and run through this cleaner today, but the file's own metadata container is not rewritten until that support lands.
What is a SynthID-class watermark, and why can't it be removed?
SynthID and similar statistical watermarks do not live in characters. They are embedded in the probabilities the model used when choosing each token — a pattern spread across word choice that no invisible-character pass can see or alter. Removing one would mean rewriting the text substantially with a different process, which changes meaning. This cleaner targets character-level and markup provenance only, so it honestly cannot affect SynthID-class signals.
Does removing watermarks make text undetectable as AI-written?
No, and the tool makes no such claim. Detectors also use stylometry and statistical signals. This is a text-hygiene utility for content you own, not a way to evade disclosure obligations.
Why does invisible Unicode break my software?
Because parsers see what renderers hide. Zero-width characters defeat string comparison, search and regular expressions; bidi controls can make source code display differently from how it compiles; no-break spaces break naive tokenisers and CSV imports.
Still stuck? Send us the codepoint.