The characters
you cannot see.

What counts as invisible

Unicode contains a family of code points that render as nothing, or nearly nothing, yet travel inside text like any other character:

• Zero-width space (U+200B) — an invisible word-break opportunity.
• Zero-width non-joiner and joiner (U+200C, U+200D) — control how neighbouring characters connect.
• Word joiner (U+2060) and the legacy byte-order mark (U+FEFF) — invisible glue.
• Soft hyphen (U+00AD) — shows up only when a word breaks across lines.
• Directional marks (U+200E, U+200F) — steer left-to-right and right-to-left rendering.
• Non-breaking space (U+00A0) — looks like a space, behaves differently.

Related but visible artefacts often travel with them: curly “smart” quotes, en and em dashes, and ellipsis characters substituted by word processors.

Where they come from

Most invisible characters arrive by accident. Word processors substitute them during autocorrect, websites embed them in templates and copy buttons, and they hitch a ride every time text is copied between applications. Some older marking tools also inserted them deliberately as a lightweight fingerprint — which is why a character-level inspection is a standard first step when examining a text’s history.

One thing they are not: Claude’s watermark. Anthropic explicitly states its announced watermark is carried by word-choice patterns and uses no hidden characters. If you find zero-width characters in a Claude draft, they came from somewhere else in the copy chain. The two concerns are separate, which is why our workspace treats them as separate tools.

What they break

Invisible characters are mostly harmless in prose but cause real failures elsewhere. They break exact-match search and deduplication, corrupt diffs and version control, make identical-looking strings compare as different in code and spreadsheets, invalidate URLs and identifiers, and trip form validation. A zero-width space inside a variable name is a genuinely miserable bug to find.

How to find them

Three reliable options, in increasing order of effort: paste the text into our Character cleanup, which counts and removes them in one pass without an LLM; switch on invisible-character rendering in an editor such as VS Code, which highlights suspicious code points inline; or search programmatically for the specific code points listed above.

How to remove them safely

Targeted removal beats blanket stripping. Delete the genuinely problematic characters — zero-width spaces, joiners, byte-order marks, directional marks that serve no purpose — while being deliberate about the rest. Non-breaking spaces are correct typography in some languages, soft hyphens may be intentional in long technical words, and joiners are meaningful in scripts such as Arabic and in emoji sequences.

For prose, a cleanup pass is safe and the visible text is unchanged. For code and URLs, review the diff before accepting changes — that is why the workspace reports the number of characters adjusted rather than silently rewriting.