Source-linked AI summary
Delving into LLM-assisted writing in biomedical publications through excess vocabulary
Dmitry Kobak, Rita González-Márquez, Emőke-Ágnes Horvát, Jan Lause
TL;DR
The paper asks how technological advances, particularly widely available LLMs, have changed biomedical scientific writing and how widespread their use is. It detects emerging LLM fingerprints directly in published abstracts and finds an unprecedented shift toward LLM-preferred style words, indicating substantial and likely underestimated usage.
Problem
The paper investigates how technological advances, particularly widely available LLMs, affect scientific writing amid concerns about inaccurate content and misuse.
Method
The study systematically detects emerging LLM fingerprints through excess-word analysis of published biomedical abstracts, avoiding reliance on separate human-written and LLM-generated text corpora.
Results
At least 13.5% of 2024 PubMed abstracts were processed with LLMs, while hundreds of LLM-preferred style words abruptly increased and some subcorpora exceeded 30% lower-bound estimates.
Takeaways & Limitations
LLMs have had an unprecedented effect on biomedical scientific writing, surpassing vocabulary changes induced by the Covid-19 pandemic and providing a measurement method for policy debates.
Takeaways & Limitations
The estimate is a lower bound because abstracts without the detected marker words are excluded, and linguistic background or publication timing may affect detectability.
Abstract
from arXiv · showhide
Large language models (LLMs) like ChatGPT can generate and revise text with human-level performance. These models come with clear limitations: they can produce inaccurate information, reinforce existing biases, and be easily misused. Yet, many scientists use them for their scholarly writing. But how wide-spread is such LLM usage in the academic literature? To answer this question for the field of biomedical research, we present an unbiased, large-scale approach: we study vocabulary changes in over 15 million biomedical abstracts from 2010--2024 indexed by PubMed, and show how the appearance of LLMs led to an abrupt increase in the frequency of certain style words. This excess word analysis suggests that at least 13.5% of 2024 abstracts were processed with LLMs. This lower bound differed across disciplines, countries, and journals, reaching 40% for some subcorpora. We show that LLMs have had an unprecedented impact on scientific writing in biomedical research, surpassing the effect of major world events such as the Covid pandemic.
1 Introduction
LLMs became widely available for scholarly writing after ChatGPT’s release, prompting concerns about research integrity and motivating efforts to measure their footprint. Existing approaches rely on labeled human- and LLM-written texts, whereas this paper proposes an unlabeled excess-vocabulary analysis of PubMed abstracts.
- Motivation: ChatGPT’s November 2022 release made human-like LLM text generation and revision widely available, including for academic writing.Researchers have integrated LLMs into daily writing and co-authored papers with them.
- Motivation: LLM-assisted scholarly writing raised concerns about factual mistakes, research integrity, bias, and misuse.The introduction specifically notes possible misuse by paper mills and inaccurate generated content.
- Research gap: Prior methods for quantifying LLM use generally assume that LLM-written text differs from human-written text and fall into three methodological groups.These include black-box detectors and approaches modeling word-frequency mixtures in scientific corpora.
- Research gap: All existing approaches share a limitation: they require a ground-truth training set of human-written and LLM-generated texts.Such datasets typically use pre-LLM human texts and prompted model outputs, introducing assumptions about models and prompting.
- Approach: The paper tracks post-ChatGPT excess word usage against pre-LLM years across more than 15 million PubMed biomedical abstracts without a labeled corpus.The approach adapts excess-mortality analysis to changes in scientific vocabulary and compares LLM-related changes with historical writing shifts.
2 Results
Excess vocabulary increased sharply in 2024, shifting from pandemic-related content terms to LLM-associated style words. Combining these markers estimated that at least 13.5% of 2024 PubMed abstracts underwent LLM processing, with substantially higher lower bounds in some fields, countries, and journals.
- Corpus and measures: 15.1 million English-language PubMed abstracts from 2010–2024 were analyzed through word-occurrence frequencies, focusing on 26.7 thousand sufficiently frequent words.The corpus was represented as a 15.1 M × 273 K sparse binary matrix.
- Corpus and measures: 2024 excess-word detection compared empirical frequency p with a counterfactual frequency q extrapolated from 2021–2022, using frequency gap δ = p − q and ratio r = p/q.The 2023 data were excluded from the extrapolation because they might already reflect LLM usage.
- Excess words: 28.0 was the 2024 excess frequency ratio for delves, while common words reached frequency gaps of δ = 0.052 for potential and δ = 0.041 for findings.These examples illustrate strong excess usage among both less common and more common words.
- Excess words: 454 excess words appeared in 2024, compared with up to 190 during the Covid pandemic, roughly one year after ChatGPT’s release.The 2024 increase followed the earlier pandemic-related rise.
- Excess words: 2024 excess vocabulary consisted predominantly of style words, contrasting with the mainly content-word vocabulary associated with the Covid pandemic.Among 379 excess 2024 style words, 66% were verbs and 14% were adjectives.
- Subcorpora: 0.25 was the estimated lower bound for Sensors, compared with 0.07 for Nature, Science, and Cell and 0.10 for Nature family journals.The authors report lower easily detectable LLM usage in journals with higher perceived prestige.
- Subcorpora: 0.41 was the lower-bound estimate for computation papers from China, while computation and bioinformatics broadly showed ∆≈0.20.South Korean papers in Sensors reached ∆ = 0.34, and some local clusters approached ∆≈0.50.
3 Discussion
LLM-assisted writing has produced an unprecedented shift toward preferred stylistic vocabulary in biomedical abstracts, with measurable but heterogeneous evidence of widespread use. The approach also has important scope limits: it detects corpus-level lexical changes rather than individual abstracts or uniquely identifying their causes.
- Limitations: Corpus-level analysis cannot identify which individual abstracts were processed by an LLM.The method estimates aggregate usage rather than assigning LLM processing to specific publications.
- Estimated prevalence: At least 13.5% of 2024 PubMed abstracts were estimated to have been processed with LLMs.Because abstracts without the marker words do not contribute to the estimate, the true fraction is likely higher.
- Heterogeneity: Estimated lower bounds ranged from below 5% to over 40% across PubMed-indexed fields, countries, and journals.Computational fields had approximately 20% lower bounds, while some subcorpora exceeded 40%.
- Limitations: Unexpected lexical changes cannot be uniquely attributed to direct LLM use because multiple topics, writing styles, or borrowed LLM-preferred words may produce similar shifts.The approach also cannot distinguish the influence of different LLMs.
- Unprecedented vocabulary change: 379 style words showed highly elevated frequencies in 2024, unlike earlier shifts dominated by content-related nouns.The excess words were mainly style-affecting verbs and adjectives preferred by LLMs.
- Methodological contribution: The approach avoids labeled human-versus-LLM training corpora and places LLM-induced writing changes in historical context.It detects emerging fingerprints directly from published abstracts and enables comparison with earlier vocabulary shifts, including Covid-related changes.
4 Materials and Methods
The study analyzes cleaned English-language PubMed abstracts from 2010–2024, computes yearly word-occurrence frequencies, and identifies excess words using conservative counterfactual projections. It also annotates excess words, analyzes selected subgroups, and documents data and code availability.
- Corpus construction: 15,103,888 complete English-language abstracts from 2010–2024 were selected from the annual PubMed snapshot after filtering abstracts to 250–4,000 characters.The source snapshot contained 24,814,136 abstracts before restricting publication years.
- Corpus construction: 286,744 abstracts were cleaned of contaminating strings, and 3,514 correction, erratum, or retraction notices were removed.Cleaning targeted journal- and year-specific strings such as editor communications, copyright notices, and citation instructions.
- Word processing: 15,103,888 × 362,441 sparse binary word-occurrence matrices were generated, with subsequent analysis restricted to 273,112 alphabetic words of at least four letters.Different strings such as mask and masks remained distinct words.
- Frequency calculation: p = (a + 1)/(b + 1) defined each yearly word frequency, where a counts abstracts containing the word and b counts all abstracts that year.The additive-one adjustment avoids division by zero.
- Excess-word analysis: For year Y, words required frequencies above 10^-4 in both Y and Y−1; q = p−2 + 2·max{p−2 − p−3, 0} produced conservative excess measures r = p/q and δ = p − q.The projection uses frequencies from Y−3 and Y−2, ensuring q is at least p−2.
- Subgroup analyses and reproducibility: Subgroup analyses covered 50 countries, 100 journals, 39 fields, inferred first- and last-author genders, and pooled journal groups, with Δ values assigned only to groups containing at least 300 papers per year from 2018–2023.The study also reports that no LLMs were used for writing or editing and makes its code, matrices, annotations, and source data available.
Supplementary Figures
The supplementary figures track excess-word patterns across successive years, summarize lemma-level counts and low-frequency 2024 style words, and visualize abstracts by field and affiliation country.
- Yearly excess words: Excess words are shown for 2013 and 2014 in Figure S1, with explanations deferred to Figure 2.The figure is part of the year-by-year supplementary analysis.
- Yearly excess words: Excess words are shown for 2015–2017 in Figure S2; mitogenome is omitted from Figure 3a because its use was concentrated in one journal.The authors considered that concentration unrepresentative of the entire biomedical literature.
- Yearly excess words: Excess words are shown for 2018–2020 in Figure S3, with explanations deferred to Figure 2.This extends the supplementary year-by-year display.
- Yearly excess words: Excess words are shown for 2021–2023 in Figure S4, with explanations deferred to Figure 2.This covers the years immediately preceding the 2024 analysis.
- Lemma counts: Figure S5 counts unique lemmas of excess words rather than individual word forms, counting inflections such as delve, delves, delving, and delved once.The same principle applies to pairs such as mask and masks.
- 2024 style words: Figure S6 displays all 291 excess style words in 2024 whose frequency is below 0.02.These words are associated with the rare-style-word analysis.
- Abstract visualization: Figure S7 visualizes 4,109,080 abstracts from 2022–2024 in two dimensions after PubMedBERT produced 768-dimensional representations and t-SNE reduced them.The visualization is colored by field and affiliation country, among other views described in the figure.