The count unit must match what people think they are counting
Aliases: grapheme versus code point · word count vs character · byte limit
What it is
The screen says “12 left.” People interpret that in their own unit: Han characters, English words or letters, bytes for engineers, septets for an SMS gateway. If the interface decrements in another unit, count-unit mismatch appears: 12 seems to remain, yet pasting one Han character or one emoji subtracts 2, or fills the cap at once. The issue is what is being subtracted, not when the counter lights, and not whether overflow is clipped.
Why it happens
“Character” in natural language is not an engineering quantity. In Unicode one visual character may be several code points (combining marks, emoji sequences); in UTF-8 Han is three bytes and emoji more; some products count weighted units. Users see glyphs; the system subtracts code points, bytes, or weights. When the step per key disagrees with expectation, the counter looks broken, or people avoid punctuation and emoji near the cap. Mixed scripts scramble further: does a space count, is an English word 1 or its letters? Unit mismatch turns “how much is left” from a brake into a riddle.
Studying it
Ask people to fill to a known cap and compare their predicted remainder with the system’s, logging disputes at emoji, combining marks, newlines, and mixed CJK/Latin. Factors: the displayed unit name (characters / letters / words / bytes), whether newlines count, whether emoji are weighted. Outcomes: prediction error, switches to “safe” characters near the cap. Do not test only BMP Han. The cheapest audit is to line up the unit word in the copy with the actual algorithm.
Where it stops holding
Storage may count bytes while the UI counts glyphs; the units can differ if the hint says so (“display characters, about N”) rather than pretending they are one number. Search keywords and tags often count tokens, not characters; moving a character counter there will lie. Unspaced languages have no intuitive “word”; an English product’s word count is meaningless for Chinese users. A screen reader that says “twelve remaining” with no unit mismatches on the auditory channel too.
Applying it
- Pick the unit that matches the task (on a Chinese UI, usually one visible character) and lock that word in the copy; do not show a bare number.
- Make one visible character (including common emoji) decrement once; if the engine must count bytes, say “in bytes; Han is about triple.”
- Audit whether newlines, spaces, and zero-width characters count, and make the copy agree.
- Verify with 10 Han characters, 10 letters, 10 emoji, and a mixed passage. If the step jumps, or the copy says “characters” while bytes are subtracted, change the algorithm or the word.