Remove invisible characters
Paste text and it names what is hidden in there, by category and count, then hands back a clean copy.
Nothing you type here leaves your device. There is no upload and no request.
Files carry their own
This handles text you paste in. Images and documents keep their hidden data somewhere else entirely — in the file’s metadata rather than in its characters — including the C2PA provenance record that AI image tools sign into what they generate. The metadata tools here strip that, and the pixels come through untouched.
Which invisible characters get removed, and which become a space
The text is walked one code point at a time and each character is checked against eight categories. The distinction that matters is what happens next. Zero-width characters, soft hyphens, bidirectional controls and tag characters are deleted outright, because they occupy no width and removing them changes nothing you can see. The unusual spaces — no-break, narrow no-break, en, em, thin, ideographic — are replaced with an ordinary space instead of deleted, because they are holding a real gap between two words and dropping them would glue those words together. That second case is the one careless cleaners get wrong, and the damage is invisible: you end up with text that reads slightly wrong and no way to tell why.
What it cannot do
- It works on text you paste in. Files carry their own metadata, which the metadata tools here handle.
- It reports what a character is, not how it got there. A zero-width space in pasted text is far more often a word processor artefact than anything deliberate.
- Formatting that depends on these characters goes with them. If you are using a no-break space to keep a number next to its unit, that break becomes possible again.
Questions people actually ask
- Is the ChatGPT watermark real?
- The characters are real; the watermark story is not. In April 2025 people noticed U+202F, a narrow no-break space, appearing where an ordinary space belonged in GPT-o3 output, and it spread as proof of a hidden watermark. Forensic analysis and OpenAI both put it down to how the model was trained — it had learned the character from high-quality typography — rather than a deliberate mark. The characters are genuinely there and genuinely worth removing, just not for that reason.
- Will this remove Claude's watermark?
- Not from text, and no character-level tool can. Anthropic's own documentation describes the text mark as woven into the text itself rather than added as hidden characters. Files are a different matter: Claude signs images it generates with C2PA provenance metadata, and Anthropic states plainly that the mark disappears when metadata is stripped. That part is removable, on the metadata tools here.
- Why does my pasted text break search or code?
- Almost always a no-break space, U+00A0, which looks exactly like a space and is not one. String comparison fails, a CSV column splits in the wrong place, a compiler rejects a line that looks fine. Word processors and web pages produce them constantly.
