How to remove invisible & zero-width characters
Some text carries characters you can’t see — zero-width spaces, joiners, byte-order marks, bidirectional controls, tag characters and look-alike glyphs. They’re used as watermarks and trackers, and they survive copy-paste. Here’s how to find and remove them, free, with Unmark.
Updated .
Which characters are “invisible”?
These are legitimate Unicode code points that render with no width or mimic normal ones. The most common culprits:
- Zero-width family — zero-width space (U+200B), non-joiner (U+200C), joiner (U+200D), word joiner (U+2060), byte-order mark (U+FEFF).
- Formatting controls — soft hyphen (U+00AD), invisible math operators, and deprecated format characters.
- Bidirectional controls — direction-flipping overrides (U+202A–U+202E) and isolates.
- Tag characters — the U+E00xx block, an invisible copy of ASCII.
- Exotic spaces & homoglyphs — no-break, thin and ideographic spaces, plus Cyrillic/fullwidth look-alikes.
Why they’re in your text
Invisible characters are inserted deliberately — as a watermark or a way to fingerprint who leaked a document — and accidentally, by copy-pasting from web pages, PDFs and word processors. Either way they can carry a hidden signal, so it’s worth stripping them before you publish or share text.
Are invisible characters the AI watermark?
For most providers, no. Google’s SynthID, OpenAI’s textGrain and Anthropic’s Claude watermark are all statistical: they shape which words the model picks, and Anthropic states plainly that its watermark adds no hidden characters. So a clean character scan does not mean text is watermark-free, and stripping characters alone will not remove a statistical mark — that takes a rewrite.
Microsoft’s Windows AI watermark is made of characters
The exception is Microsoft. Starting in October 2026, text from the Windows AI text APIs — the Phi Silica generate, rewrite and summarise calls — can include an invisible watermark made of four code points in a row: U+200B (zero-width space), U+034F (combining grapheme joiner), U+2060 (word joiner) and U+200B again. Microsoft inserts it zero, one or several times per response, and the positions change.
Because it is built from ordinary characters, it is fully detectable and fully removable: the checker flags the exact sequence, and Unmark strips every copy. Microsoft notes that apps may have their own obligations to mark AI-generated content, so if you are building on these APIs, check those first. Details and other providers are in the AI watermark tracker.
Invisible characters are a separate fingerprint. They turn up in AI output when text passes through chat interfaces, editors and copy-paste pipelines — narrow no-break spaces and zero-width characters are common — and some tracking schemes insert them deliberately to tell copies of a document apart. They are easy for software to detect, which makes them worth removing even when no statistical watermark is involved.
How to remove them
- Paste your text into Unmark. It scans instantly, in your browser — nothing is uploaded.
- Reveal the hidden characters to see exactly what’s there, colour-coded by type.
- Strip them and copy the clean text. Load-bearing marks (like emoji sequences) are preserved, so nothing visible changes.
This removal is deterministic and complete — unlike statistical watermarks, invisible characters can be removed with certainty.
Free, 100% in your browser.