What counts as a hidden character
| Group | Examples | Why it appears |
|---|---|---|
| Zero-width | U+200B zero width space, U+200D joiner, U+2060 word joiner, U+FEFF BOM | Fingerprinting/tracking markers, file artifacts, emoji internals, spam-filter evasion |
| Invisible formatting | U+00AD soft hyphen, U+034F combining grapheme joiner | Word processors, hidden markers |
| Look-alike spaces | U+00A0 no-break space, U+2009 thin space, U+202F narrow NBSP | Typographic output of editors, PDFs and chat interfaces |
| Directional controls | U+202E right-to-left override, isolates | Legitimate RTL text — or visual spoofing (“Trojan Source”) |
| Tag characters | U+E0001–U+E007F | Invisible ASCII copies — a documented channel for smuggling hidden text and prompt-injection payloads |
Common questions
Are hidden characters proof of AI?
No. Some AI interfaces emit special spaces, but so do Word, Google Docs, iOS keyboards and every second website. Treat hidden characters as a fact about the text, not a verdict about its author — that distinction is the core of our methodology.
Can removing them strip an AI watermark?
Not a real one. Modern AI text watermarks (Claude, SynthID text) are statistical patterns in word choice, not characters — see how AI text watermarks work. The clean-up tool here exists to fix broken copy-paste artifacts and remove tracking markers, and it tells you exactly what it removed.
Why do positions matter?
Regularly spaced invisible characters suggest deliberate marking; a single stray NBSP suggests an editor artifact. The checker reports positions so you can judge the pattern yourself.