PaperlayerView exact revision on arXiv

Source-linked AI summary

An LLM-Associated Register Shift in Korean Journal Abstracts: A Morphology-Aware Excess-Vocabulary Study, 2018-2026

Aron Lee

arXiv:2609.07447v1cs.CLcs.DL

TL;DR

Research on LLM-associated vocabulary has focused on scholarly English, leaving open whether similar shifts occur in structurally different scholarly languages. This paper adapts excess-vocabulary analysis to Korean and Vietnamese abstracts and finds a Korean shift beginning in late 2024, with conditional bounds rising through 2026.

  • Problem

    Prior excess-vocabulary studies focus on English, leaving limited evidence about whether the method and LLM-associated shifts apply to structurally different scholarly languages.

    Source: §1 · p. 1

  • Method

    The study adapts excess-vocabulary analysis to morphological units in Korean and segmented units in Vietnamese, using KCI and VJOL abstracts with placebo and model-generated controls.

    Source: §1 · p. 1, §3.1 · p. 3, §4.7 · p. 4

  • Results

    33.0% is the 2026 split-half set-based conditional lower bound for Korean abstracts, following a late-2024 onset and a rise through 2025 that flattens in mid-2026.

    Source: §1 · p. 2, §8 · p. 23

  • Takeaways & Limitations

    The findings support a Korean scholarly register shift that is not confined to articles whose English abstracts show model vocabulary, while the estimated bounds remain conditional rather than prevalence estimates.

    Source: §1 · p. 2, §8 · pp. 23–24

    The Korean analysis depends on one morphological analyzer and selected fusion rules, which affect which words are named even though switching off adjectival fusion leaves the bound unchanged.

    Source: §7 · p. 23

Abstract

from arXiv · show

Excess vocabulary, a word's frequency above its pre-2023 trend, is how the change in scholarly English after 2022 has been measured. We adapt it to Korean with morphological units on 398,296 KCI abstracts (2018-August 2026), with 47,165 Vietnamese abstracts for comparison. Placebo floors are 0.1-2.2 points for the single-word statistic and at most 2.9 for the re-selected split-half set statistic. Korean abstracts show nothing in 2023, onset in late 2024, a rise through 2025 flattening in mid-2026: sisahada "suggest" appears in 21.4% of 2026 abstracts against 5.3% expected; plain verbs like araboda "look into" fall to a quarter of trend. Under stated assumptions the single-word conditional lower bound on LLM-processed abstracts is 3.5%, 10.5% and 16.1% for 2024-2026 and a split-half set bound 7.8%, 20.6% and 33.0%. Holzwarth et al.'s estimator under the same discipline gives 41.9% and 72.1% for 2025-2026. Subject-matter controls reduce but do not remove it: restricting the set to lemmas three language-model annotators all call style leaves 14.7 of the 33.0 points, and pairing each 2026 abstract with its journal's closest base-period abstract leaves 34.1. Tested translation routes do not explain it: the surface marks of translated Korean fall as the markers rise. In the same articles' English abstracts the excess appears a year earlier; where the English side carries none, the Korean shift persists at 30 to 66% of the rate where it does. Control abstracts from three providers reproduce the rising words, with marker turnover consistent with model generations; implied prevalences are scenario-dependent.

1 Introduction

Prior work measures LLM-associated vocabulary shifts in scholarly English, leaving open whether comparable signals appear in structurally different scholarly languages. This study adapts excess-vocabulary analysis to Korean and Vietnamese and finds a late-2024 Korean shift that persists across robustness checks and comparisons.

Research gap

  • Earlier studies document English lexical shifts, but Korean scholarly evidence ended in 2024 and Vietnamese scholarly writing had not been studied with this approach.Existing work also leaves open whether whitespace-token methods transfer to languages with different word structures.Source: §1 · p. 1

Contributions

  • The study analyzes 398,296 Korean KCI abstracts through August 2026 and uses 47,165 Vietnamese abstracts as an external comparison, not for a prevalence bound.It adapts excess vocabulary to morphological units and adds a journal panel and structured-abstract normalization to address composition artefacts.Source: §1 · p. 1

Validation and alternative explanations

  • Placebo tests put false-positive levels at 0.1–2.2 percentage points for the single-word statistic and −0.8 to +2.9 points for the split-half set statistic.The positive control tests whether models reproduce high-excess markers, while style-or-topic and translation checks examine alternative explanations.Source: §1 · pp. 1–2

Within-article comparison and robustness

  • The English abstracts of the same articles show excess a year before Korean, but the Korean shift also appears when the English abstract carries no English markers.Robustness checks find the rising and falling vocabularies negatively associated within abstracts, a pattern consistent with a document-level register shift.Source: §1 · p. 2

Findings

  • Korean’s single-word conditional lower bound on LLM-processed abstracts rises from 3.5% in 2024 to 10.5% in 2025 and 16.1% in 2026, while the split-half set bound reaches 33.0% in 2026.The signal begins in late 2024, rises through 2025, and flattens in mid-2026; sisahada appears in 21.4% of 2026 abstracts against 5.3% expected.Source: §1 · p. 2

Resources

  • The paper releases per-year document-frequency tables, analysis code, and a Korean AI style dictionary for reuse without re-harvesting the corpora.Source: §1 · p. 2

2 Related work

Corpus-level studies estimate LLM-associated writing from vocabulary distributions, while research on why models favor particular words and how those preferences cross languages motivates this study’s focus. Prior Korean evidence is limited, and human diffusion may also contribute to observed lexical shifts.

2.1 Corpus-level measurement of LLM-assisted writing

  • Corpus-level methods infer LLM-associated prevalence from distributional shifts without labeling individual documents, and excess vocabulary can provide a lower bound from surplus word occurrences.The approach requires a pre-LLM series for extrapolation and does not require an LLM reference distribution.Source: §2.1 · p. 2

2.2 Why models prefer particular words

  • Preference-tuning rather than pre-training data is proposed as the likelier source of some favored words, raising the question whether other languages show translated favorites or formal-register equivalents.Source: §2.2 · p. 2

2.3 Beyond English

  • Prior Korean scholarly work found modest increases in Korean abstracts through 2024, while the Vietnamese scholarly literature lacked a published distributional study.The current study extends the Korean series, changes the units to lemma–tag forms with suffix fusion, and adds Vietnamese as an external comparison.Source: §2.3 · p. 2

2.4 Diffusion into human language

  • Excess vocabulary in abstracts can reflect both model-assisted text and people whose vocabulary has shifted, so the lower bound does not separate those processes.A study of spontaneous English speech found abrupt rises in ChatGPT-preferred words after release [12].Source: §2.4 · p. 2

3 Data

The paper analyzes Korean and Vietnamese scholarly abstracts and uses generated Korean abstracts as positive controls. The datasets cover different periods and scopes, and abstract-based measurements do not represent full-text prevalence.

3.1 Korean: KCI abstracts

  • The completed KCI corpus contains 398,296 unique Korean abstracts from 2,282 journals, published from 2018 through August 2026.The analysis caps each year at a common size rather than comparing raw counts across years.Source: §3.1 · p. 3
  • KCI journals over-represent established Korean-language humanities, social-science, education, and applied-science journals, while abstracts are more polished and more likely to be model-edited than full text.Source: §3.1 · p. 3

3.2 Vietnamese: VJOL abstracts

  • The Vietnamese dataset contains 47,165 abstracts analyzed from 2019–2025; the changing set of contributing journals is addressed with a journal panel.Source: §3.2 · p. 3

3.3 Positive-control corpus

  • The positive-control corpus includes generated, edited, rewritten, and translated Korean abstracts based on pre-ChatGPT 2019 material, with prompts and lengths varied across conditions.The drafting arm’s length effect is not fully separated from its different title sample.Source: §3.3 · p. 3

3.4 Licensing and release

  • The release provides per-year frequency tables, journal-panel lists, code, and generated positive-control abstracts, but does not redistribute harvested abstracts.Source: §3.4 · p. 3

4 Methods

The methods adapt vocabulary tracking to Korean and Vietnamese units, control for corpus composition and formatting, and estimate excess against extrapolated pre-LLM trends. Bounds are interpreted conditionally, with placebo and positive controls used to assess their behavior.

4.1 Lexical units

  • Korean lexical units are lemma–part-of-speech forms with selected derivational suffixes fused; style words include non-noun units and qualifying bigrams, while plain common nouns count as content.Vietnamese uses segmented word–part-of-speech units, with verbs, adjectives, adverbs, and bigrams treated as style words.Source: §4.1 · p. 3

4.2 Structured-abstract normalisation

  • A regular expression removes structured-abstract section labels at clause starts before tokenization, preventing formatting changes from appearing as lexical excess.Source: §4.2 · pp. 3–4

4.3 Journal panel and capping

  • The loose journal panel requires at least five abstracts in both base and target periods, and annual samples are capped so no year dominates the frequency floor.A strict every-year panel was also tested and gave the same ranking with smaller samples.Source: §4.3 · p. 4

4.4 Excess frequency

  • Expected frequency is a least-squares extrapolation from base years; the excess ratio denominator is floored to prevent near-absent units from generating arbitrarily large ratios.The study uses five base years and compares alternative expectation rules and extrapolation horizons.Source: §4.4 · p. 4

4.5 Uncertainty

  • Uncertainty intervals use journal-cluster bootstrap resampling, but intervals for units selected on the same data do not have nominal 95% coverage.The split-half set bound chooses markers on one journal half and measures them on the other; marker selection predates the final run but used an earlier pass over the same corpus.Source: §4.5 · p. 4

4.6 Controls

  • The negative control treats 2022 as a pre-LLM target, while the positive control compares generated-marker frequencies with real abstracts and, for editing, with originals.Source: §4.6 · p. 4

4.7 Lower bound on LLM-processed abstracts

  • The maximum excess frequency across style words is a lower bound on the processed share because each abstract can contribute at most one occurrence of a unit.The study reports single-word and split-half set bounds, subtracts matched-horizon placebo levels, and treats LLM attribution as conditional on assumptions that exclude non-model sources of excess.Source: §4.7 · p. 4

4.8 Fields and multi-word constructions

  • KCI field groups are assigned by journal-name keyword rules, and non-adjacent multi-word constructions are counted with regular expressions on raw text.Source: §4.8 · p. 4

5 Results

The results trace a Korean abstract register shift emerging in late 2024 and continuing through 2026, alongside marker patterns in model-written abstracts, across fields and Vietnamese abstracts, and in English abstracts of the same articles. They assess the trend’s robustness and conditional prevalence estimates while examining placebo floors, translation markers, and an alternative estimator.

5.2 Time course: a lag year, an onset, and acceleration

  • After no 2023 movement, Korean marker vocabulary rose from 2024; in 2026 시사하다 appeared in 21.4% of abstracts versus 5.3% expected.Plain verbs also fell: 알아보다 reached 1.3% in 2026, or 0.24 of its expected 5.7%.Source: §5.2 · pp. 5–6, Figure 1 · p. 6

5.3 The dictionary: what rose and what fell

  • In 2026, 시사하다 had a +16.1 pp excess, and 19 of 65 tracked markers exceeded 5 pp of excess.The rising list also featured 확인되다 (+14.1 pp) and 규명하다 (+11.6 pp); the largest 2022 placebo correction was 0.5 pp, leaving the ordering unchanged.Source: Table 3 · p. 6, §5.3 · p. 7, Table 4 · p. 7

5.4 Constructions

  • The construction 단순한 X를 넘어 rose to 4.93% of abstracts in 2026, 61 times its trend expectation.X을 시사한다 closed 20.8% of 2026 abstracts, while 이러한 결과 opened 10.6%.Source: §5.4 · p. 8, Table 5 · p. 8

5.5 Fields

  • The rising-marker index roughly doubled to quadrupled across fields from 2019 to 2026, with 2026 shares ranging from 46.9% in engineering and IT to 71.8% in business and economics.No field moved before 2023, and every field turned from 2024; medical and natural-science Korean abstracts are not representative of all work in those fields.Source: §5.5 · pp. 8–9, Table 6 · p. 9, Figure 4 · p. 9

5.6 Positive control: the models write the rising words

  • In the length-matched drafting control, models used many rising markers at rates close to or above real 2026 abstracts, but produced 구조적 in only 2–6%.For example, the model used 시사하다 in 27% of abstracts versus 22% in real 2026 text; falling markers such as 알아보다 were rarely used.Source: §5.6 · pp. 9–10, Table 7 · p. 10, Figure 5 · p. 12
  • Across model conditions, marker turnover tracked model generation more than provider: 2024-vintage models reproduced the receding 2024 group, while 2026 models favored newer markers.Light editing moved vocabulary less than drafting, and the tested models also reproduced several markers through rewriting; structural remained poorly captured across conditions.Source: §5.6 · pp. 9–10

5.7 External comparison: Vietnamese

  • Vietnamese register markers appeared above trend in 2025, a year after Korean and two years after English, but none cleared the 2022 placebo floor in percentage points.The largest 2025 excess was 3.7 pp for nhấn mạnh, below the 9.1 pp control maximum, so the study reports no Vietnamese prevalence bound.Source: §5.7 · p. 12, Table 8 · p. 12

5.8 Robustness

  • Pre-LLM placebo targets produced maximum style-word excesses of 0.7–1.5 pp, placing the 2023 value of 1.0 pp inside the control band.The maxima were 1.5 pp for 2020 and 2021 targets and 0.7 pp for 2022; the study does not interpret 2023 as a shift.Source: Table 2 · p. 5, §5.8 · p. 13
  • The ratio-restricted single-word bounds were 3.5, 10.5, and 16.1 pp in 2024–2026, while density normalization lowered the 2026 bound to 13.8 pp.The 2026 estimate remained 7.3 times the largest measured four-year single-word placebo floor; placebo floors grew with extrapolation horizon, reaching 2.2 pp for the ratio-restricted statistic at four years.Source: §5.8 · p. 13
  • The rise spans most journals and survives composition, base-year, selection, and suffix-fusion checks.Among 482 journals, 시사하다 rose in 91%; equal journal weighting changed rising-marker excess ratios by less than 11%. Disabling suffix fusion left the bound unchanged, and selecting and measuring markers on different journal halves gave 16.3 pp for 2026.Source: §5.8 · pp. 13–14
  • Monthly 시사하다 usage turned upward in late 2024, rose through 2025, and flattened around 20–23% in 2026; January–July equal-month values were similar.It reached 23.4% in April 2026 and then remained between 20 and 23% through July; the 2026 January–July value was 21.0% against 21.5%.Source: §5.8 · p. 13
  • In 2026, abstracts with rising markers were less likely to contain falling markers, unlike in 2019 and 2022.For example, 살펴보다 occurred in 7.3% of 2026 abstracts with a rising marker versus 12.7% of other abstracts; the pattern is consistent with whole-document register rewriting but does not distinguish machine from human rewriting.Source: §5.8 · p. 14

5.9 Lower bound on LLM-processed abstracts

  • The single-word lower bound rose from 0.1 pp in 2023 to 3.5, 10.5, and 16.1 pp in 2024–2026; the split-half set bound reached 33.0 pp in 2026.The 2026 single-word marker 시사하다 appeared in 21.4% of abstracts against 5.3% expected; the 47-lemma set appeared in 60.1% against 27.0% expected. The set bounds were 7.8 pp in 2024 and 20.6 pp in 2025; no set passed selection in 2023.Source: §5.9 · pp. 14–15, Table 9 · p. 15, Figure 7 · p. 16

5.10 Sensitivity, specificity and the scenarios the floor allows

  • Matched-length drafting scenarios imply that 77–97% of 2026 abstracts were processed, but these are scenario values, not prevalence estimates.Under polishing scenarios, the observed broad-set vocabulary cannot be reproduced at any prevalence; the processing mix and model sensitivities are unknown, so only the 33.0% lower bound survives without a sensitivity assumption.Source: §5.10 · pp. 16–17, Table 13 · p. 17

5.11 Lexical diversification

  • Mean distinct lemmas per abstract rose from 87.2–87.8 in 2018–2023 to 97.8 in 2026, despite only modest length growth.Among 600–800-character abstracts, the mean rose from 89.1 lemmas in 2022 to 98.0 in 2026.Source: §5.11 · p. 17
  • The 10th, 50th, and 90th percentiles of lexical density each rise 1.0–1.1 per 100 characters from 2022 to 2026, with standard deviation essentially unchanged.Polishing real abstracts leaves density unchanged, while drafting produces denser text and rewriting raises density by 0.71 (OpenAI) and 0.89 (Claude); the observed corpus-wide rise is consistent with drafting or heavier rewriting, not the tested polishing condition.Source: §5.11 · p. 17

5.12 The English abstracts of the same articles

  • English abstracts of the same articles show excess vocabulary in 2023, about a year before the Korean shift begins in late 2024.The English single-word bound is 13.2 pp in 2024, 23.1 in 2025, and 34.3 in 2026, versus 3.5, 10.5, and 16.1 for Korean.Source: §5.12 · pp. 17–18, Figure 8 · p. 19
  • 30 to 66% of the Korean marker rate remains among articles whose English abstract carries no period-matched English markers, under the stated sensitivity assumptions.The odds ratio between the English and Korean marker flags rises to 2.2 in 2025 and 2.3 in 2026, from a pre-2023 baseline of 1.6 to 1.9; the sensitivity correction assumes conditional independence.Source: §5.12 · p. 18, Table 15 · p. 18, Figure 8 · p. 19

5.14 Translation as a route

  • Translation markers fall as Korean marker vocabulary rises; model translation produces 에 의해 in 11.1% of output, versus 5.7% in the 2026 corpus.The surface features are informative in direction but are not a translation detector; translation remains possible for individual articles, but the evidence does not support it as the general account of 2026 English vocabulary.Source: §5.14 · p. 20, Table 18 · p. 20, Table 19 · p. 20

5.15 An alternative estimator

  • 9.3%, 41.9%, and 72.1% are the cross-fitted Korean estimates for 2024–2026 under the reimplemented alternative estimator [2].Choosing thresholds on the measurement half instead yields 9.3%, 48.6%, and 81.1%. The 2025 and 2026 cross-fitted values exceed the estimator's placebo range, but the 2024 figure does not; comparison with English biomedical estimates is not controlled.Source: §5.15 · p. 21

6 Discussion

The Korean shift is concentrated in formal academic language, but its vocabulary changes alongside model generations and does not identify who produced any given abstract. The reported bounds are conditional on an attribution assumption; Vietnamese evidence is later and less conclusive, while the tables are intended for calibration rather than individual-paper detection.

6.1 A formal-register signature, not a translated one

  • Korean's rising academic verbs, evaluative adjectives, and discourse frames differ in form from English excess words, while plain native expressions decline.The twelve Sino-Korean verbs rise from 0.40 expected occurrences per abstract in 2022 to 1.11 in 2026; ten native expressions fall from 1.07–1.12 through 2022 to 0.80 in 2026.Source: §6.1 · p. 22
  • The Korean academic increase and Juzek's Korean news-word decline can coexist because the studies differ in register, timing, and text-production setting.The news series ends in 2024, before the Korean academic onset in late 2024; the authors also note that polishing alone installs the rising words in their control.Source: §6.1 · p. 22
  • In the same articles, English markers rise in 2023 before Korean markers begin in late 2024, and the Korean shift persists where English carries none.After correcting for the English indicator's sensitivity, the Korean rate in that group is 30–66% of the rate where English carries markers.Source: §6.1 · p. 22

6.2 Marker turnover is consistent with model-generation turnover

  • Marker turnover tracks model generation in the tested conditions, but the evidence is consistent with—not proof of—a generation effect.The paper cautions that the receding group was selected for its 2024 excess and the positive control uses five models and two prompt styles.Source: §6.2 · p. 22

6.3 What the lower bound does and does not say

  • Under A2, the 2026 lower bounds are 16% by the single-word statistic and 33% by the set statistic; without A2, they bound an unattributed register change.The bounds include abstracts polished by a model and cannot separate direct model use from human writers who have adopted the style.Source: §6.3 · pp. 22–23
  • For 2026, drafting scenarios imply 77–97% processed abstracts, but polishing scenarios cannot reproduce the observed vocabulary at any prevalence.These scenario conversions depend on the assumed prompt and model; the authors report the lower bound rather than replacing it with a prevalence estimate.Source: §6.3 · p. 23

6.4 Vietnamese: the same words, one year later

  • Vietnamese shows a small, late pattern: no excess beyond the negative-control level through 2024, followed by several rising words in 2025.The authors cannot distinguish later chatbot adoption, weaker earlier Vietnamese generation, and the smaller corpus as explanations.Source: §6.4 · p. 23

6.5 Implications for editors, reviewers and teachers

  • The tables are for calibration, not individual-paper detection; teachers may use the falling list to identify ordinary Korean expressions students are losing.The authors also present the lower bound as a defensible policy-document figure in place of English-based prevalence numbers.Source: §6.5 · p. 23

7 Limitations

The study measures abstracts from a large but non-exhaustive journal sample, and its 2026 data cover only January through August. Several comparisons and prevalence conversions also depend on measurement choices, model conditions, or assumptions that limit how broadly their estimates can be interpreted.

7 Limitations

  • The analysis covers abstracts rather than full text, a non-exhaustive KCI harvest, and January–August 2026; publication year also lags writing behavior.VJOL covers a fraction of Vietnamese journals, with coverage growing over time.Source: §7 · p. 23
  • Field assignment uses journal-name keyword rules, placing about 30% of panel journals in “other” and assigning interdisciplinary journals by their first matching keyword.Year sizes differed in the first harvest; the corrected corpus moved the 2025 single-word bound up 1.1 pp and set bounds up by as much as 1.9 pp.Source: §7 · p. 23
  • Korean units depend on one morphological analyser and chosen fusion rules, while the Vietnamese tagger is less accurate and its segmentation was not manually verified.Turning adjectival fusion off leaves the bound unchanged but changes which words are named; another analyser or fusion choice could reclassify units.Source: §7 · p. 23
  • Model-based calibration uses a limited set of providers and prompts, while the field's model mix and declared-use ground truth are unknown.The implied shares are conditional on the assumed processing; short-draft conditions are downward-biased and are not comparable across conditions.Source: §7 · p. 23
  • The cross-language comparison cannot separate shared model authorship from translation, and its English abstracts are 79% complete.The English marker set uses word forms rather than lemmas; its sensitivity correction supports the direction of the comparison, but not a precise size.Source: §7 · p. 23
  • The marker list was not preregistered, style/topic labels have only fair model-annotator agreement (κ = 0.24), and the design cannot exclude all convention-driven change.The translation features are surface marks, not a detector, and cannot rule out careful translations that leave none of those marks.Source: §7 · p. 23

8 Conclusion

Using morphological units, the study finds a rising Korean academic-register signature, with lower bounds on abstracts carrying excess vocabulary and a later, less conclusive Vietnamese pattern. Robustness checks and same-article comparisons support the persistence of the shift, but the bounds remain conditional and do not identify direct model authorship.

8 Conclusion

  • The Korean signature begins in late 2024 and reaches 16.1% by the single-word lower bound and 33.0% by the set bound in 2026.The set-based bound is 7.8% in 2024 and 20.6% in 2025; placebo levels are at most 2.2 points for the single-word statistic and 2.9 for the set statistic.Source: §8 · p. 23
  • Under the same split-half discipline, the Holzwarth estimator gives 41.9% for 2025 and 72.1% for 2026.When its threshold is chosen on the measurement half, it gives 48.6% and 81.1%, respectively.Source: §8 · p. 24
  • The shift survives multiple robustness checks, and appears in nine of every ten journals with enough abstracts to assess.The checks include alternative expectation models, journal weighting and panels, month-composition matching, and removal of the fusion rule.Source: §8 · p. 24
  • Subject-matter and journal-matched comparisons leave 14.7 of 33.0 points and 34.1 points, respectively; tested translation marks fall as the style markers rise.The conclusion says the tested subject-matter controls and translation routes do not explain most of the shift.Source: §8 · p. 24
  • English abstracts of the same articles shift about a year earlier than Korean, while Vietnamese shows a similar family of words a year later.Where English carries no markers, the corrected Korean rate remains 30–66% of the rate where English does.Source: §8 · p. 24

9 Declaration of language-model use

The author says language models were used heavily in producing the paper, while retaining responsibility for the question, analysis direction, verification demands, and claims. The declaration distinguishes those writing and coding roles from models used as annotators or experimental subjects.

9 Declaration of language-model use

  • The author conceived and directed the study and is solely responsible for its claims, while the agent wrote and executed the analysis code and drafted the English manuscript.The author states that they wrote none of the code or English, but supplied the Korean instructions and approved every version.Source: §9 · p. 24
  • The paper used models both as writing aids and as experimental subjects, but marker-style classification was the only measurement in which a model acted as judge.Control abstracts, sensitivity conditions, and translation conditions were model outputs whose reported numbers depend on those outputs.Source: §9 · p. 24
  • The declaration reports 100% model-agent production of analysis code, build and gate scripts, and English text, with tracked values checked against result files.Table 20 lists the models by role; the agent's underlying models changed with availability.Source: §9 · p. 24, Table 20 · p. 24

Appendix J

Appendix J distinguishes reproducible corpus arithmetic from interpretive prose and documents the manuscript’s own marker-count check. It also notes that lexical counts do not capture sentence shape or rhythm.

Appendix J

  • The deposited code rebuilds every table and figure from frequency tables and result files, while estimator choices and interpretation remain outside that archive.The appendix specifically flags readings of annotator agreement and topic-matched shifts as judgments not recoverable from the archive.Source: Appendix J · p. 24
  • The manuscript contains 37 occurrences of the fixed 33-word English marker list in 24,846 words, or 14.9 per ten thousand.The comparison rates are 5.7 for KCI English abstracts from 2018–2022 and 58.5 for model-written English abstracts.Source: Appendix J · p. 24
  • All 37 marker occurrences are mentions, quotations, listings, reported frequencies, or Vietnamese glosses, rather than markers used in their own right.The count uses the fixed 33-word list; the appendix notes that another 47-word set includes ordinary words used in the manuscript itself.Source: Appendix J · p. 24
  • The lexical statistic is blind to sentence shape and rhythm, limiting the corpus results as well as the manuscript’s self-count.The appendix reports Claude Sonnet 5, GPT-5.6 Sol, and EXAONE 3.5 7.8B Instruct among the listed tools or models.Source: Appendix J · pp. 24–25

Data availability

The archive provides corpus statistics, panel lists, generated controls, analysis code, and figure and table generators. Raw abstracts are not redistributed, but can be re-harvested from public endpoints using the included scripts.

Data availability

  • Per-year frequency tables, journal panels, field rules, generated controls, harvesting and analysis code, and figure and table generators are archived and downloadable.The passage identifies Zenodo and a downloadable reproducibility package.Source: Data availability · p. 25
  • Raw abstracts are not redistributed; the included scripts can re-harvest them from the public KCI and VJOL OAI-PMH endpoints.Source: Data availability · p. 25

Appendix A · Korean marker words with bootstrap intervals

Appendix A reports bootstrap intervals for Korean marker units, while warning that selection and subsampling distinguish these estimates from the corpus-wide values quoted in the text. The listed units include both rising and falling markers.

Appendix A · Korean marker words with bootstrap intervals

  • Table A1 gives point estimates and 95% intervals from 1,000 longitudinal journal-cluster bootstrap replicates, but its units were selected on an earlier pass over the same corpus.It uses at most 12,000 abstracts per year, so estimates can differ from corpus-wide Tables 3 and 4, which use up to 40,000 abstracts per year.Source: Table A1 · p. 25
  • 탐구하다 (“explore”) has a 2024 ratio of 1.90 [1.60, 2.22] and 2025 excess of +2.1 [1.5, 2.7].Source: Appendix A · pp. 25–26
  • 살펴보다 (“take a look at”) has 2026 excess of −14.2 pp [−16.0, −12.4], among the listed falling marker words.Source: Appendix A · p. 26

Appendix B · Lexical density

The appendices document lexical-density handling, structured-abstract label removal, and keyword-based journal field assignment. Density normalization reduces some excess bounds, while the auxiliary rules specify filtering and classification procedures.

Appendix B · Lexical density

  • Density normalization reduces the 2026 bound from 16.1 to 13.8 pp and the unrestricted 2025 bound from 11.0 to 9.3, without changing ordering or conclusions.Distinct units are counted on a fixed 12,000-abstract sample per year; filtered document-frequency totals also vary with yearly sample size.Source: Appendix B · p. 26, Table A2 · p. 26, Table A3 · p. 26

Appendix C · Structured-abstract label pattern

  • The pre-tokenization pattern removes a structured-abstract label only when it begins a clause and is followed by a colon.This leaves the same words untouched when they occur inside sentences.Source: Appendix C · p. 26

Appendix D · Field rules

  • Journals are assigned to the first field group whose keyword list matches the journal name; unmatched journals are classified as Other.Source: Appendix D · pp. 26–27

Appendix E · Suffix fusion removed

Removing noun–suffix fusion leaves the single-word bound unchanged but changes which units populate the marker list; 구조적’s rise survives as the bare noun 구조.

Appendix E · Suffix fusion removed

  • Fusion prevents 적, 화, and 성 suffixes from being dropped and counted only as part of a content noun.Removing fusion replaces eleven fused units among the thirty largest 2026 excesses with eleven other style units; the other nineteen keep almost the same order.Source: Appendix E · p. 27, Table A4 · p. 27
  • The bound is unchanged to the first decimal with fusion removed, because no fused unit attains it.The bound is attained by a verb or adjective in every year.Source: Table A4 · p. 27, Appendix E · p. 27
  • 구조적’s rise survives removal of fusion as a rise in the bare noun 구조, so it is not an artefact of the rule.It is the one rising marker the positive control does not reproduce, but the bound does not rest on it.Source: Appendix E · p. 27

Appendix F · Selection and estimation on different journals

Selecting markers in one half of the journal panel and measuring them in the other yields bounds close to the same-half estimates, with 2023 at the placebo level.

Appendix F · Selection and estimation on different journals

  • Journals are hash-split into halves: markers are ranked on half A, then the largest excess among its twenty selected markers is measured on untouched half B.The table’s naive columns select and estimate on the same half.Source: Appendix F · p. 27, Table A5 · p. 27
  • 16.3 pp in 2026, 11.2 pp in 2025, and 5.0 pp in 2024 are the out-of-sample bounds, matching same-half values to one decimal.In 2023, the out-of-sample value is 1.0 pp, at that horizon’s placebo level.Source: Appendix F · p. 27, Table A5 · p. 27
  • Reversing which half selects and measures gives 15.9 pp for the 2026 single-word statistic, 38.1 pp for the rare set, and 24.1 pp for the top-k set.Across 100 journal splits, the median 2026 set bound is 32.4 pp, with a 5th–95th percentile range of 31.2 to 33.8.Source: Appendix F · p. 27

Appendix G · Base-year sensitivity

Refitting the trend without the weakest base year, 2021, barely moves the bounds; the word attaining the bound remains 시사하다.

Appendix G · Base-year sensitivity

  • The sensitivity check addresses 2021’s initially thin, skewed base-year sample: 12,975 abstracts, 94% published in January–March.The completed corpus held 37,665 panel abstracts for 2021.Source: Appendix G · p. 27
  • 16.1 to 16.4 pp in 2026 and 11.0 to 10.9 pp in 2025 are the bounds when 2021 is dropped.The word attaining the bound is 시사하다 in all three fits.Source: Appendix G · p. 27, Table A6 · p. 27
  • Dropping both 2021 and 2018 gives bounds of 16.3 pp for 2026 and 11.0 pp for 2025.The table reports that the 2026 bound moves by at most 0.3 pp when 2021 is excluded.Source: Appendix G · p. 27, Table A6 · p. 27

Appendix H · Re-analysis on the completed harvest

Reanalysis on the completed harvest leaves the 2026 bounds well above the extended placebo floors. The appendix also reports how set composition, lemma-only placebos, and polishing affect the set indicator.

Appendix H · Re-analysis on the completed harvest

  • The second harvest completes 2021 and adds 2017–2018, while 2015–2016 were not systematically harvested; analysis years are capped at 40,000 abstracts.The completed-corpus reanalysis keeps the pipeline otherwise unchanged and extends the placebo horizon to four years ahead.Source: Appendix H · pp. 27–28, Table A7 · p. 28
  • Completing 2021 changes the 2025 bound from 9.4 to 10.5 pp and lowers the 2024 bound by 0.2 pp; 2021’s month composition does not carry the result.The 2026 excess ratio of 시사하다 changes from 4.13 to 4.04.Source: Appendix H · p. 28
  • 16.1 pp for both single-word statistics and 33.0 pp for the set are the 2026 bounds, against extended placebo floors of 2.2 pp and 2.9 pp, respectively.Completing the harvest moves the 2026 bound from 16.2 to 16.1 pp; adding 2017 as a sixth base year gives 16.4 pp.Source: Appendix H · p. 28

Appendix I · Set bound details

  • Lemma-only placebo values range from −1.5 to +2.9 pp across the grid, generally below all-unit placebo values.At four years ahead from two base years, the rare-lemma set gives −1.2 pp and the top-k set 0.0 pp.Source: Appendix I · p. 28, Table A10 · p. 29
  • The 47 style lemmas selected on half A occur in 60.1% of 2026 half-B abstracts against 27.0% expected.The set-bound gap peaks at k = 5 for top-k sets and T = 1% for rare sets, with larger or more common units diluting the gap.Source: Table A8 · p. 28, Table A9 · p. 28
  • Polishing inserts rare style lemmas into roughly a fifth to a third of abstracts that lacked them, while removing them from roughly half that had them.Consequently, polishing produces only a small net rise in the set indicator.Source: Table A11 · p. 29

Appendix J · Version history

The version history records eight numbered versions and a corrected copy of version 5.1, tracing changes from the first complete draft to the current version. Revisions document expanded analyses, softened claims, and reproducibility and disclosure updates.

Appendix J · Version history

  • The history records eight numbered versions and a corrected version 5.1, which removes duplicated text and updates data-availability counts without changing claims or results.Version 5 added same-article English-abstract analyses; the corrections were carried into version 6.Source: Appendix J · p. 29
  • The first complete draft analyzed KCI and VJOL corpora; version 2 adds the split-half set bound, controls, journal-cluster intervals, and completed-harvest analyses.Version 2 also adds the headline single-word statistic restricted to units with excess ratio at least 1.5.Source: Appendix J · p. 29
  • Versions 3 and 4 add set-statistic placebo and implied-share analyses, soften causal wording, and restate the within-document result as a strong negative association.Version 4 also adds a rewriting condition to the positive control and notes provider variance in falling vocabulary.Source: Appendix J · p. 29
  • Version 6 states identification assumptions A1–A3, labels implied shares as scenarios, and adds style/topic, translation-route, and English-abstract analyses.The analysis is recomputed on the completed harvest, with a manifest check comparing key counts, bounds, set sizes, and DOIs.Source: Appendix J · p. 29
  • Version 7 completes the reproducibility pipeline, adds bootstrap and split-half resampling details, and replaces drifted prose values with machine-filled results.It also marks placebo multiples as descriptive, not test statistics, and identifies marker annotators as language models at first mention.Source: Appendix J · p. 29
  • Version 8 adds a full account of language-model use and reports a reimplementation of [2] beside, rather than instead of, the conditional lower bound.The version history says the split-half estimate is lower by 6.7 and 9.0 points and is the figure the paper stands behind.Source: Appendix J · p. 30
Loading 2609.07447v1…