Korean journal abstracts changed in late 2024
We adapted excess-vocabulary measurement to Korean morphology, ran it over 398,296 abstracts, and found a register shift that starts in late 2024 and flattens in mid-2026.
Excess vocabulary is a simple idea: take a word's frequency before a change, extrapolate the trend, and read off how far the observed frequency sits above it. It was used to show how language models moved scholarly English. Nobody had run it on Korean, and the reason is mechanical rather than a lack of interest.
Korean does not hand you words
English excess vocabulary counts whitespace-delimited words. Korean does not work that way. 구조적 is a noun plus a derivational suffix, and a tokenizer that splits them puts the signal on the noun, which is a topic word, not a style word. We fuse the XSN class back onto its stem before counting. Turning that rule off does not change the bound we report; it only changes which word gets the credit, and it moves it onto vocabulary that is about the subject rather than the register.
The corpus
398,296 Korean abstracts from KCI, 2018 through August 2026, plus 47,165 Vietnamese abstracts as an exploratory comparison. Journals are the unit of clustering, not documents, because abstracts inside a journal are correlated and treating them as independent narrows the intervals by about 40%.
What the series looks like
Nothing in 2023. Onset in late 2024. A rise through 2025 that flattens in mid-2026. 시사하다 ("suggest") appears in 21.4% of 2026 abstracts against 5.3% expected from its own pre-2023 trend. Plain verbs move the other way: 알아보다 ("look into") falls to a quarter of trend.
The numbers we are willing to defend
Under stated identification assumptions, the single-word conditional lower bound on LLM-processed abstracts is 3.5% in 2024, 10.5% in 2025 and 16.1% in 2026. A split-half set statistic, which selects marker words on one half of the journals and measures on the other, gives 7.8%, 20.6% and 33.0%.
These are conditional lower bounds, not prevalence. They say how much cannot be explained away under the assumptions we state, and the assumptions are in the paper where they can be attacked.
What we checked before believing it
Subject matter is the obvious confound: fields grow and shrink, and their vocabulary moves with them. Restricting the marker set to lemmas that three language-model annotators all independently call style rather than content leaves 14.7 points. Pairing each 2026 abstract with the closest abstract from its own journal in the base period leaves the difference at 34.1.
Translation is the other obvious route, since a Korean abstract is often written after an English one. The surface marks of translated Korean fall over the same period the markers rise. In the same articles' English abstracts the excess appears a year earlier, and where the English side carries no markers at all the Korean shift persists at 30 to 66% of its uncorrected rate.
Paper
An LLM-Associated Register Shift in Korean Journal Abstracts, INTFRAME Research, DOI 10.5281/zenodo.22102389. CC BY 4.0, reproduction package included. Seven numbered versions are public, and the earlier ones are still readable, including the claims that did not survive review.