Hidden Unicode vs Statistical Text Watermarks

The practical differences matter more than the terminology.

• Made of: literal invisible code points, versus a pattern in which words were chosen.
• Findable by: scanning for characters, versus a keyed statistical test you generally cannot run yourself.
• Removed by: deleting the characters, exactly, versus rewriting the wording, partially.
• Survives copy-paste: usually yes, but easily lost to reformatting, versus yes, because it is the wording.
• Tells you: nothing reliable about origin, versus that a keyed model was likely involved.
• Provable claim: an exact count of what was removed, versus a reduction that cannot be measured without the detector.

Zero-width spaces, word joiners, soft hyphens, bidirectional controls and non-breaking spaces all exist for legitimate typographic reasons, documented in the Unicode standard. Web pages use them for line-breaking control, editors insert them during formatting, PDF exports scatter them, and messaging clients add them around emoji and mentions.

So when a blog post presents a zero-width space as proof of AI authorship, the reasoning is broken twice over: the character has ordinary explanations, and the providers that do watermark text say they add no characters at all.

Invisible characters cause real, unglamorous problems. Two visually identical strings stop being equal, so search misses matches, diffs show phantom changes, form validation rejects correct input, identifiers and URLs break, and copied code fails to compile with no visible cause.

The Unicode Consortium's own security guidance covers this class of confusion. Cleaning is worth doing as formatting hygiene — a claim we can make exactly, with counts, and without any reference to watermarks.

Neither mechanism establishes authorship. A character proves nothing about who inserted it. A statistical watermark indicates a keyed model was involved somewhere — Anthropic notes it cannot separate writing from heavy editing — and says nothing about ownership.

Any tool, or any accusation, that treats either signal as proof of who wrote something is overreaching.