Since 2023 several studies have shown that the vocabulary of English scholarly writing changed after ChatGPT. Kobak et al. (Science Advances, 2025) called it excess vocabulary: a word's frequency above the trend it was on before 2023. English abstracts started overusing delve, underscore, intricate.
One Korean study exists: Koo, Kim and Kim (Language and Information, 2026) ran an excess-vocabulary comparison over ~880,000 KCI humanities and social-science abstracts, projecting 2024 from 2021-2022. They found a sharp rise in the English abstracts and only modest movement in the Korean ones, and their observation ends in 2024 — before the Korean onset we measure. Korean had not been measured at the point where it actually moves, and the method does not transfer for free: Korean is agglutinative, so whitespace tokens are the wrong unit, and the verbs that carry register are built by fusing a Sino-Korean noun with a verbalising suffix.
So we rebuilt the statistic over morphological lemma-tag units and ran it on the full openly harvestable record of the Korea Citation Index: 398,296 abstracts from 2,282 journals, 2018 to August 2026, with 47,165 Vietnamese abstracts as an exploratory second case.
What the numbers say
The verb 시사하다 (sisahada, "suggest, imply") appears in 21.4% of 2026 abstracts. Extrapolating its 2018–2022 trend predicts 5.3%.
In the other direction, plain-register expressions fell: 알아보다 ("look into") and 도움이 되다 ("be helpful") dropped to about a quarter of their trend.
The time course is sharp. Nothing in 2023. Onset in the second half of 2024. A steep rise through 2025. Flattening in mid-2026.
Under stated identification assumptions, the conditional lower bound on the share of LLM-processed Korean abstracts is:
| Statistic | 2024 | 2025 | 2026 |
|---|---|---|---|
| Single word | 3.5% | 10.5% | 16.1% |
| Split-half word set | 7.8% | 20.6% | 33.0% (95% CI 30.8–35.4) |
This is not a detector
It cannot tell you whether your paper was written by a model, and it is not built to. It measures a distribution over a corpus and reports a floor: the true share is at least this, under assumptions we state and test.
Ruling things out
A number this large invites obvious objections, so we tried to kill it several ways.
"The fields just changed what they write about." We paired every 2026 abstract with the base-period abstract from the same journal whose title is most similar (character-bigram cosine, IDF-weighted). The matched difference is 34.1 pp against 34.0 pp unmatched. Topic matching barely moves it. We also had three independent language-model annotators classify each marker lemma as style or topic; restricting the set to lemmas all three call style still leaves 14.7 pp.
"It's translation from English." The surface marks of translated Korean (the 것이다 construction, ~에 대한, explicit plural 들) fell over the same period while the markers rose.
"It's a statistical artefact of extrapolation." We ran the identical procedure with pre-ChatGPT target years, at every extrapolation horizon the data allow. The largest set-statistic value any placebo produced was 2.9 pp. The 2026 value is more than eleven times that.
"You got lucky with one split." The set is chosen on one half of the journals and measured on the other. We repeated the split 100 times with different journal hashes: median 32.4 pp, 5th–95th percentile 31.2–33.8, set sizes 38–56 lemmas.
The English abstracts of the same articles
Most KCI articles carry an English abstract alongside the Korean one. That gives an article-fixed comparison: same authors, same journal, same paper, two languages.
The English side moves a year earlier — the excess is already there in 2023. And among articles whose English abstract carries none of the period's English markers, the Korean shift is still present, at 30–66% of its uncorrected rate after correcting for the English indicator's own sensitivity.
So the Korean change is not simply a by-product of a model-written English abstract, though the size of the effect does shrink under that correction.
Everything reproduces
The paper (28 pages), the analysis code, per-year document-frequency tables for Korean, English and Vietnamese, the journal panels, and 1,437 control abstracts we generated with models from three providers are all deposited:
https://doi.org/10.5281/zenodo.22110398
Unzip it into an empty directory, run reproduce.sh, and every table and figure in the paper is regenerated from the shipped results. It ends in REPRODUCE_OK or it fails loudly. Raw abstract texts are not redistributed for copyright reasons, but the harvesting scripts against the public KCI and VJOL OAI-PMH endpoints are included.
Two free tools built from the same measurements, in Korean:
- AI style dictionary — which words rose and fell, with base rates
- Style checker — paste text, see which markers it contains (again: a base-rate lookup, not a verdict)
What this does not settle
The attribution assumption is the weak point, and we say so in the paper. Observational data cannot rule out that some contemporaneous stylistic drift, entirely unrelated to language models, produced part of this. What we can do is show the size of the change, its timing, its resistance to topic and translation controls, and the level of noise the method produces when nothing is happening. That is why the headline number is a conditional lower bound, not a prevalence.
The annotation of markers as style-versus-topic was itself done by language models, which is circular in a way we would rather fix with human annotators. That is the first thing on the list.
The manuscript is under review at Scientometrics. Comments, replications and objections are welcome — the data is right there.
Top comments (0)