The tells turn over
The words that marked model-written Korean in 2024 are falling in 2026 while a different set rises. Anything built on a fixed list of tells has a half-life.
The useful part of the Korean abstract study, for anyone who is not a linguist, is not the headline number. It is that the marker set is not stable.
What we saw
We generated control abstracts with three providers under matched conditions and compared their vocabulary against the corpus. The controls reproduce the rising words, and they show marker turnover that tracks generation rather than provider. Words that were characteristic in 2024 are on the way down in 2026, and a different set is on the way up.
That pattern is visible in the corpus too, and it has an innocent competing explanation we say out loud in the paper: words selected because they were extreme in one year will regress toward the mean in the next, and observation alone cannot separate that from real turnover. We report the turnover as consistent with the controls, not as established by them.
Why it matters outside linguistics
Any tool that decides whether text was model-written from a fixed list of words is dated the day it ships. It does not fail loudly. It keeps returning confident answers about a vocabulary that has moved on, and the failure mode is that it gets quieter, not wronger-looking.
The same applies to the reverse use, editing to avoid the tells. The list you edit against is last year's list, and the words you replace them with are on somebody's next list.
The signal that does not depend on a word list
The measurement we trust most is not a frequency count at all. It asks, within a single abstract, whether the presence of a rising word predicts the absence of a falling one. In 2019 and 2022 it does not: the two vocabularies coexist as you would expect from independent drift. In 2026 they become mutually exclusive.
That is a different kind of evidence. Gradual adoption of a fashionable word does not push its old synonym out of the same document. A rewriting pass does. And the test does not care which words are fashionable this year, only that a substitution happened.
What we published, and what not to do with it
The marker dictionary and a checker are public alongside the paper, and the reproduction package has the frequency tables. Use them to see how a corpus moved.
Do not use them on a person. The bounds in the paper are conditional lower bounds on a population under stated assumptions, and the implied prevalences are scenario values, not an estimate. Nothing in this work licenses a verdict about one document, and a tool that turns a population statistic into an accusation is a misuse of it regardless of how careful the statistic was.
An LLM-Associated Register Shift in Korean Journal Abstracts, DOI 10.5281/zenodo.22102389, CC BY 4.0.