Short answer: paste the text. The table tells you how many of each kind were found and removed; the output is the text without them.
What gets removed
- Zero-width — U+200B to U+200D, U+2060, U+FEFF.
- Bidi controls — U+202A to U+202E and U+2066 to U+2069.
- Soft hyphens — invisible break hints from PDF extraction.
- Control characters — no printable form, usually export damage.
- Variation selectors — emoji presentation markers.
The honest limit
Text produced by a language model can carry a statistical watermark — a bias in which words were chosen, invisible to a reader and to this page. It is not a character, so no character-level tool can find or remove it, and this page will never tell you a document is “AI-free”. What it can do is show you exactly which invisible characters were in the text and take them out.
Cleaning a document before comparing it? Paste both versions intotext diff afterwards.