How an invisible Unicode watermark appears
Digital text is a sequence of encoded characters, not merely the shapes you see. Some Unicode characters affect boundaries, joining, direction, or rendering without drawing a visible glyph. Copy and paste usually carries those characters even when the destination looks unchanged. That makes them useful for typography and language support, but it also lets software hide identifiers or payloads in otherwise ordinary prose.
A reliable scanner examines code points rather than pixels. It can distinguish an ordinary space from a zero-width space and show a directional isolate that a text area does not reveal. The evidence is literal and reproducible. The interpretation still needs context: the same code point can be an accidental copy artifact, intentional formatting, part of a writing system, or a deliberate marker.
Zero-width spaces and no-break controls
U+200B ZERO WIDTH SPACE creates a word boundary without a visible gap. It can help scripts whose words are not separated by ordinary spaces, and it sometimes appears after copying from layout software or websites. U+FEFF began as a byte-order mark and can also appear inside text as a zero-width no-break space. In a plain English paragraph, a stray occurrence may be safely removable, but a scanner should still show its location.
U+2060 WORD JOINER does the opposite of a break opportunity: it asks layout not to wrap at that point. U+00AD SOFT HYPHEN marks a possible hyphenation point and may become visible only at a line break. Removing these characters can alter typography even when the current viewport looks unchanged. NoLLMWM reports them and leaves the context-sensitive pair unselected by default.
Joiners, variation selectors, and emoji
U+200D ZERO WIDTH JOINER combines characters into one displayed sequence. It is central to many emoji involving families, professions, gender presentation, and skin tones, and it participates in shaping for multiple scripts. U+200C ZERO WIDTH NON-JOINER prevents joining where characters would otherwise connect. Removing either with a broad regular expression can split an emoji or change the written form of a word.
Variation selectors such as U+FE0F request emoji-style presentation, while supplementary variation selectors support other standardized glyph choices. Some subdivision flags use an emoji base followed by invisible tag characters and a terminator. A cleanup that strips every tag would break those flags. NoLLMWM recognizes and preserves these common contexts rather than presenting every invisible scalar as disposable.
Bidirectional controls need special care
Arabic, Hebrew, Persian, and other right-to-left writing often appears beside left-to-right numbers, URLs, or interface labels. Unicode bidirectional marks, embeddings, isolates, and pop controls help the rendering algorithm keep these fragments in the intended order. They are invisible but semantically important to display. Removing them can produce confusing or misleading visual order while the underlying characters remain unchanged.
Direction overrides can also be abused to disguise filenames or source code, which is why security tools flag them. The correct response is inspection, not automatic deletion. NoLLMWM names all supported bidi controls, explains that they change direction, and never preselects them. A person familiar with the language and expected rendering should decide whether a mark is required, accidental, or malicious.
A safe cleanup procedure
Work on a copy. Run a scan and record every code point, location, and warning. First remove only clearly accidental characters in a language context you understand. Rescan, then visually compare line breaks, emoji, mixed-direction fragments, search behavior, copied identifiers, and code. If the text will enter a database or publishing system, test that destination because normalization and rendering can differ.
Avoid one-line recipes that remove the entire ranges U+200B–U+200F or U+202A–U+202E. Those ranges combine spaces, joiners, marks, and direction controls with distinct roles. Avoid Unicode normalization as a universal remedy; normalization combines certain equivalent sequences but does not simply erase all controls, and it can change identifiers. Selective, explained transformations are safer and auditable.
- Preserve the source before any deletion.
- Default to retaining joiners, variation selectors, and bidi controls.
- Review multilingual text with a competent reader.
- Scan the final exact string after every copy or export step.
What Unicode cleanup says about watermarks
Removing an identified physical payload can remove that payload from the new string. It still does not prove who inserted it or whether another watermark mechanism exists. Statistical token signals reside in word choices; document metadata resides in a container; C2PA assertions may be signed; and image watermarks can live in pixels. A plain-text Unicode result must not be generalized to those categories.
The most trustworthy report names the exact change: for example, “removed two U+200B characters from an authorized copy and preserved all reported joiners.” That is stronger than saying “all AI watermarks removed,” because another person can reproduce it. NoLLMWM is designed to make the narrower statement easy.