Attach a chat export. Paste a new message. The page reads each person's writing habits — compression residue, character n-grams, ㅋ density, sentence endings — and names its suspect.
runs 100% in your browser · nothing uploaded| Name | Messages | Corpus | Data |
|---|
Accuracy expectations come from a 72M-character Korean SNS corpus, measured for the scorer this page actually uses. Two things dominate: how many people are in the room, and how much you paste. More corpus keeps helping — there is no point where it stops mattering. Your group may differ, and the Yes/No button is how this page learns its real hit rate.
A write-up of how this was built and, mostly, of how its accuracy figure was wrong three separate times: a leak worth up to 71 points when the pasted message is already in the uploaded file, a parser hole that put the sender's name into the query and inverted the evidence, and a list of thirteen things I got wrong. English only.
How It Works →