https://orcid.org/0009-0005-7029-0108
Figures
Abstract
Large Language Models (LLMs) have ushered in a transformative era in Natural Language Processing (NLP), reshaping research and extending NLP’s influence to other fields of study. However, there is little to no work examining the degree to which LLMs influence other research fields. This work empirically and systematically examines the influence and use of LLMs in fields beyond NLP. We curate 106 LLMs and analyze ∼148k papers citing LLMs to quantify their influence and reveal trends in their usage patterns. Our analysis reveals not only the increasing prevalence of LLMs in non-CS fields but also the disparities in their usage, with some fields utilizing them more frequently than others since 2018, notably Linguistics and Engineering together accounting for ∼45% of LLM citations. Our findings further indicate that most of these fields predominantly employ task-agnostic LLMs, proficient in zero or few-shot learning without requiring further fine-tuning, to address their domain-specific problems. This study sheds light on the cross-disciplinary impact of NLP through LLMs, providing a better understanding of the opportunities and challenges.
Citation: Pramanick A, Hou Y, Mohammad SM, Gurevych I (2026) Transforming scholarly landscapes: The influence of large language models on academic fields beyond computer science. PLoS One 21(1): e0337127. https://doi.org/10.1371/journal.pone.0337127
Editor: Bekalu Tadesse Moges, Woldia University, ETHIOPIA
Received: July 2, 2025; Accepted: November 4, 2025; Published: January 14, 2026
Copyright: © 2026 Pramanick et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All code, scripts for dataset download, and results files used in this study are publicly available at: https://github.com/UKPLab/plosone2025-llm-trends. The dataset used in this study is separately accessible at: https://github.com/allenai/s2orc?tab=readme-ov-file.
Funding: This work has been funded by the German Research Foundation (Deutsche Forschungsgemeinschaft, DFG) as part of the Research Training Group KRITIS (No. GRK 2222). This work has also been funded by the LOEWE Distinguished Chair “Ubiquitous Knowledge Processing”, LOEWE initiative, Hesse, Germany (Grant Number: LOEWE/4a//519/05/00.002(0002)/81), and by the “Open-Access-Publikationskosten” programme of the German Research Foundation (Deutsche Forschungsgemeinschaft, DFG). The funders had no role in the study design, data collection and analysis, the decision to publish, or the preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Modern science heavily relies on citing past research to build upon past ideas, situate the proposed work, reject old hypotheses, etc., thereby facilitating the dissemination of good ideas [1]. Some ideas are limited in the scope of their influence (being cited narrowly within a specific subfield). At the same time, others may have broad applicability, influencing not only an entire field such as Computer Science (CS) but also many disciplines beyond. Since the number of published papers is too large for manual review [2,3], there is growing work in automatically and quantitatively tracking the influence of ideas within and across fields [4,5].
Arguably, one of the most transformative ideas over the past decade is that of LLMs [6]. Originally proposed within NLP, LLMs have revolutionized almost all research areas within NLP itself [7,8]. Moreover, their utilization is not confined to NLP; other fields are leveraging LLMs as well [9,10]. While it is well recognized that these models are being adopted outside of NLP, the full extent and nature of their usage remains unclear. Analyzing the widespread adoption and utilization of LLMs across various fields provides essential insights for promoting responsible AI practices.
Although recent studies have begun to explore the influence of LLMs within CS [11,12], their impact in other disciplines beyond CS is still largely unknown. While it is evident that non-CS fields are paying attention to LLMs [13], precise details about which fields are leveraging them and the specific purposes of their use are still to be understood.
The modes of influence are multifaceted and intricate in nature, which makes it challenging to empirically determine the degree of influence. In this work, we focus on a specific dimension of influence: the scientific impact that one field has on another [14]. One notable marker of this inter-field influence is citation [15]. Therefore, we propose that the degree to which a source field cites the works of a target field can serve as a rough indicator of their influence [16]. Although citation patterns are subject to biases, meaningful aggregate-level insights can still be obtained [17,18].
While no universally accepted definition of LLMs exists [7], in this work, we take LLMs as foundational models [8] built on the transformer architecture [19], pretrained on massive textual datasets with over 100M parameters. We carefully curated a dataset of 106 well-cited LLM papers up to February 2024 including BERT [20], T5 [21], GPT-3 [22], PaLM [23], ChatGPT [24], and LLaMA [25]. We then quantitatively investigate:
- (A) Which non-CS fields are impacted by LLMs? And to what degree? (section Breadth of LLM adoption)
- (B) How do the usage patterns of LLMs evolve over time within these non-CS fields? (section Evolving usage patterns of LLMs)
We complement the above discussions with an additional qualitative analysis exploring:
- (C) In what contexts are LLMs applied within these non-CS fields? (section Applications of LLMs in non-CS fields)
In addressing the questions above, we utilize the Semantic Scholar data [26] to construct our dataset comprising ∼148k papers from 22 fields outside of CS, published between 2018 and February 2024, that cite LLMs. This dataset includes structured full texts extracted from the PDFs of the papers as well as their metadata, including the fields of study, year of publication, venue of publication, and author information. Additionally, we include similar data from the papers that originally introduced these 106 LLMs, together with ∼273k LLM citations from the aforementioned papers outside of CS.
While the quantitative analyses ( A and B) provide insights into the adoption of LLMs across diverse fields, they do not capture the nuances of how LLMs are being utilized within these fields. To gain deeper insights for C, we delve beyond metadata, exploring the content of the papers citing LLMs. Our qualitative examination of papers’ contents not only illuminates diverse applications of LLMs outside of CS but also guides towards potential improvements and research directions for better LLM usability in various non-CS fields. Finally, we approximately assess and discuss the extent to which these papers, especially in fields like Biology, Medicine, Psychology, Law, etc., discuss ethical concerns associated with LLMs, such as hallucinations [27] or non-reproducible outputs [28], using a smaller human-annotated dataset.
In summary, our findings suggest that 1) linguistics, engineering, and medicine are the top three fields citing LLMs, and fields more closely aligned with CS tend to more readily adopt LLM technology compared to other domains (section Breadth of LLM adoption); 2) BERT continues to be the preferred LLM among non-CS fields, even in terms of average yearly citations. Moreover, many of these fields frequently cite task-agnostic models like GPT-3 [22] or LLaMA [25], which excel in few-shot settings without needing additional fine-tuning (section Evolving usage patterns of LLMs); and 3) non-CS fields mostly use LLMs to solve their domain-specific problems (section Applications of LLMs in non-CS fields). Our study is the first attempt to empirically and systematically investigate the impact and usage of LLMs in fields outside CS.
The aim of our study is not to establish causal claims about the transformative impact of LLMs in non-CS fields, but rather to provide a systematic descriptive mapping of their diffusion. By presenting large-scale empirical evidence of how LLMs are cited and applied across 22 fields, we offer a baseline for future in-depth studies. Our intended audience includes NLP researchers interested in interdisciplinary exchange as well as domain specialists who may benefit from understanding how LLMs are entering their fields. We hope this work can raise awareness, identify opportunities for tailored model development, and highlight gaps in ethical awareness, thereby supporting more effective and responsible cross-disciplinary collaboration.
Related work
Scientific trends analysis
Early work by [29] served as a catalyst for exploring scientific trends in NLP. Research in scientific trends analysis primarily spans three dimensions, from citation patterns to metadata and content analysis. A vast amount of research has been done into the study of citation patterns and the amalgamation of topological measures within citation networks to assess research trends [30–32]. Complementary to this, [33] did a content analysis, employing rhetorical framing to elucidate trend patterns. [34] and [35] take a metadata-driven approach, employing prevalence correlation to examine the interplay between publication topics and research grants. [36] apply causal techniques to explore how tasks, methods, datasets, and metrics interact within the context of the evolution of NLP research. [37] explores dataset usage patterns across research communities and observes distinct trends in dataset creation within NLP communities, shedding light on the nuanced dynamics of research resource utilization. While the majority of the previous research in Computer Science and NLP examines research trends within Computer Science and NLP, respectively, in this work, we analyze both citations and content to investigate the influence of LLMs (CS and NLP technology) on fields outside of Computer Science.
NLP scientometrics
In parallel, NLP scientometrics has witnessed significant advancements in recent years as researchers strive to understand the landscape of NLP research and its evolution [38,39]. Prior studies have examined the engagement of NLP research with other disciplines, shaping the research trajectory [40,41]. While various studies have observed a rising trend in interdisciplinary interactions occurring both across fields and within subfields [42,43], this growth primarily manifests in connections between closely related domains, with limited increases in associations between traditionally distant research areas, such as materials sciences and clinical medicine [44]. We add to this line of work by investigating how research fields that were traditionally distant from NLP are now incorporating LLM techniques to address their challenges.
LLM abilities assessment
LLMs are a category of transformer-based language models distinguished by their size, typically consisting of hundreds of millions or even more parameters. They are trained on vast and diverse data [45]. These models exhibit remarkable proficiency in understanding and generating natural language, enabling them to excel in complex language-related tasks [46,47]. However, LLMs come with challenges, including the generation of plausible yet factually incorrect text [48]. Moreover, they can be manipulated to produce harmful content, raising ethical concerns about their misuse [49,50]. Our study does not assess LLM capabilities or risks but focuses on how non-CS fields use LLMs, and we examine ethical concerns related to LLMs based on the content of the papers within these fields.
Research methodology
Our study adopts a large-scale descriptive scientometric approach to explore how Large Language Models (LLMs) have diffused across research fields beyond Computer Science (CS). We focus on three overarching research questions.
- (A) Which non-CS fields are impacted by LLMs, and to what degree?
- (B) How have usage patterns of LLMs evolved over time within these fields?
- (C) In what contexts are LLMs applied within these fields, and to what extent are ethical concerns acknowledged?
To answer these questions, we detail our process of data collection, field classification, and analysis procedures.
Data collection
We began by curating a list of 106 influential LLMs up to February 2024, drawing on the Stanford Ecosystem Graphs [51]. This resource systematically tracks widely cited foundation models and their associated publications.
Choosing the LLMs.
The Computer Science community has yet to establish a universally accepted definition of LLMs. In this study, we operationalize LLMs as foundation models: transformer-based architectures pretrained on massive textual datasets, with more than 100M parameters. Proprietary LLMs without identifiable academic publication records were excluded from the bibliometric analysis, though they may still appear in qualitative analyses. This approach prioritizes models with sufficient coverage in bibliographic data to support reproducible large-scale analysis. We note that our curated set naturally emphasizes well-established, widely documented models such as BERT. In contrast, more recent instruction-tuned, multi-modal, or domain-specific LLMs were excluded because (a) they do not always have an identifiable associated publication, (b) their bibliometric coverage remains limited or inconsistent across databases, and (c) their short citation history prevents meaningful longitudinal trend analysis.
Using Semantic Scholar Dataset (S2) [52], we retrieved ∼148k papers published between 2018 and February 2024 that cite one or more of the curated LLMs. This dataset includes structured full texts (when available), titles, abstracts, author information, citation contexts, and metadata such as year, venue, and field of study. In addition, we extracted ∼273k citation instances from these papers to the curated LLM set.
Field classification
To situate LLM adoption across disciplines, we relied on Semantic Scholar’s S2FOS3 field-of-study classifier, which assigns each paper to one or more of the 22 fields with a reported accuracy of 86%. The 22 fields outside CS span natural sciences, social sciences, and the humanities. When a paper was associated with multiple fields, it contributed proportionally to each.
For author-level analyses, we identified each researcher’s primary field as the domain in which they most frequently published. This allowed us to distinguish collaboration patterns between CS and non-CS researchers. While misclassifications are possible, aggregate analyses across large datasets mitigate their impact, and robustness checks using author primary fields yielded consistent trends.
Analytical procedures
Our analyses combined quantitative citation statistics with content analysis to address each research question.
- (A) Breadth of LLM Adoption: We quantified the extent of LLM citations across non-CS fields by calculating the share of citations per field relative to the total. To measure inequality in distribution, we computed the Gini index, where values closer to 0 indicate more even diffusion. We further normalized citation counts by overall publication volume in each field to assess depth of adoption. For historical context, we compared LLM uptake with prior CS exports (Hidden Markov Models [53], Recurrent Neural Networks [54], and Long Short Term Memories [55]) to contextualize their influence.
- (B) Evolving Usage Patterns: To capture temporal trends, we analyzed year-wise adoption rates of different LLMs across fields. We measured popularity using the annual average number of citations per model and computed the Mean Age of Citation (mAoC) to estimate how quickly each field engages with newer LLMs. We also examined patterns of interdisciplinary collaboration by calculating the proportion of CS authors collaborating with non-CS researchers in LLM-citing papers. We note that this affiliation-based proxy provides only a first-order approximation of interdisciplinarity, as some non-CS authors may hold CS training or dual appointments.
- (C) Applications of LLMs in non-CS Fields: To explore contexts of use, we extracted citation contexts, titles, and abstracts mentioning LLMs. We processed these using regex-based heuristics to identify trigrams indicative of tasks (e.g., “protein function prediction”, “legal judgment prediction”). Irrelevant phrases were manually filtered. Additionally, we conducted a keyword search to estimate how frequently ethical risks (e.g., bias, hallucination, reproducibility) are explicitly discussed in these papers, complemented by a small human-annotated subset for validation.
Heuristic validation
To improve reliability, we conducted manual evaluations. For task extraction, a check of randomly sampled 100 representative trigrams confirmed that over 80% corresponded to meaningful research tasks. To evaluate the reliability of our ethical-risk detection method, we manually examined 100 randomly selected papers addressing ethical concerns related to LLMs. This analysis showed high recall, as only 7 relevant papers discussed such concerns without explicitly mentioning LLMs. We further validated precision by examining 200 randomly selected papers and found no cases where ethical issues were mentioned in the same context as LLMs without being identified by our method. While these evaluations are approximate, they suggest that the heuristics are sufficiently reliable for large-scale descriptive analysis.
Ethics statement
We clarify that the number of citations used in our analysis should not be employed to diminish any specific field or its investments based on low citation counts. Decisions in science and research should rely on a multifaceted evaluation that considers aspects like popularity, relevance, resources, impact, and geographical and temporal factors. This approach prevents oversimplified interpretations and recognizes the diversity and complexity of research fields and their contributions.
Data
We initiate our study by compiling a list of 106 LLMs from the Stanford Ecosystem Graphs [51]. This curated compilation includes the names of the LLMs selected based on their number of citations (for details, refer to S1 Appendix). We manually identified and included in our list the papers that introduced these LLMs.
Following [56], we use the Semantic Scholar (S2) dataset [52] to obtain the target papers. Specifically, we include around 148k papers from non-CS fields from S2 that cite the above-mentioned LLMs. Thus, our dataset contains structured full texts extracted from the PDFs of these papers along with their metadata, encompassing author details, publication year, publication venue, and field of study. In Table 1, we illustrate our dataset statistics.
We prefer S2 over other sources like the ArXiv dataset due to its broader coverage, including ArXiv papers. Additionally, S2 employs a field-of-study classifier (S2FOS3) based on abstracts and titles to identify the field of study (with 86% accuracy). This dataset also contains information on paper citations, including the time and venue of the first publication of the citation, citation context, and citation intents (S2 dataset overview and statistics are described in S1 Appendix and Table 2).
Further, by analyzing author names, unique author ids, and their associations with published papers in S2, we identify the field where an author has published most frequently, which we consider as their primary field of interest.
Main analyses
We use the dataset described above to examine the degree of LLM adoption in fields outside of NLP and CS, addressing three primary aspects: the scope of LLM influence (section Breadth of LLM adoption), their evolving utilization(section Evolving usage patterns of LLMs), and finally, their applications in these fields(section Applications of LLMs in non-CS fields). Subsequently, we offer a detailed examination of each facet.
Breadth of LLM adoption
To gauge the scope of LLM influence, we address three research questions: first, we investigate the extent of LLM adoption in fields beyond NLP and CS; second, we examine the depth of LLM adoption within these fields and finally, we compare LLM popularity to similar past technologies in non-CS fields.
- (Q1) How widely have LLMs permeated broader academic disciplines (beyond NLP and CS)?
Method.
We examine the number of citations from non-CS papers to our curated set of 106 LLMs. If a citing paper is labeled to be in multiple fields, then it contributes to the citation count in each field. We calculate the percentage of LLM citations attributed to each non-CS field relative to the total citations from all non-CS fields. Further, we use the Gini Index [57] on the relative citation counts to quantify the degree of deviation from a perfectly equal citation distribution across fields. A Gini index of 0 indicates a perfectly uniform distribution, while an index of 1 denotes all the probability mass concentrated on one value.
Results.
Fig 1(a) is a Sankey plot of the incoming citations from non-CS fields to LLM papers , with the width of the grey flow path representing the citation volume. Certain fields stand out in terms of LLM citations. Linguistics accounts for the maximum number of citations (23.19%), followed by Engineering (22.28%), Medicine (17.63%), Environmental Science (6.89%), and Mathematics (5.08%). Fig 2 reports the Gini index of citation distribution across fields, which has declined over time from 0.77 in 2018 to 0.65 in 2023, suggesting broader adoption.
(a) % citations from non-CS fields to LLMs (left); (b) % CS authors collaborating with non-CS fields (right).
Inequality in citation distribution across fields.
Discussion.
Initially, citations were concentrated in Linguistics and Engineering, but over time, LLM usage has spread to more fields. The declining Gini index reflects this diversification. Interestingly, the share of LLM citations by field correlates only weakly with CS collaboration (Pearson 0.28, Fig 1(b)), suggesting that adoption is driven less by interdisciplinary co-authorship and more by perceived utility within fields. To further illustrate the evolving citation trends, in S3 Appendix, we present a year-wise analysis of citations to LLM research papers from non-CS fields.
Further, to assess the depth of LLM penetration across each of these disciplines, examining the variations in publication volumes among these fields is important. In other words, a field with a high number of LLM citations may not be highly influenced by them when viewed in light of the total number of papers published in that field, and vice versa. This guides our exploration of the following section Q2.
- (Q2) To what degree have LLMs penetrated various non-CS fields?
Method.
Expanding the analysis in Q1, we now investigate the popularity of LLMs within these fields. For each non-CS field, we calculate the percentage of papers citing LLMs out of the total number of papers published in that field. This acts as a rough indicator of how extensively the field is utilizing LLMs.
Results.
Fig 3 shows the LLM citation counts relative to the field’s publication volume. Proportionally, Linguistics has the most extensive engagement, followed by Law, Mathematics, Art, and Engineering with LLMs compared to other fields. Although Medicine contributes many absolute citations, its proportion of citing papers is smaller compared to fields such as Mathematics and Environmental Science.
Discussion.
Fields with strong ties to computational methods, such as Mathematics, display relatively high proportional adoption. This reflects both the central role of LLMs as computational models and the tendency of these fields to analyze their underlying properties. By contrast, citation prevalence in Medicine is tempered by the large size of the field, illustrating that absolute citation counts alone may exaggerate impact.
We further analyze the influence of CS author collaborations on research papers within each field by calculating the average number of CS authors per paper within that field since 2018. As illustrated in Fig 4, the fields with the highest average number of CS author influence are Linguistics, Engineering, and Mathematics. We observe a positive Pearson correlation of 0.38 between the average number of CS authors per paper in a non-CS field and the percentage of papers within that field that cite LLMs (indicating moderate correlation). This empirical evidence indicates that although the total number of CS authors collaborating within a field has little consequence, an increased number of CS authors on a research paper raises the likelihood of using LLMs in that work.
- (Q3) To what extent have non-CS fields embraced LLMs compared to other technologies from CS?
Method.
We can gain a better understanding of how markedly LLMs are influencing other fields by comparing their influence to other past CS exports. To explore this, we examine the impact of two iconic technologies: Sequence Models (Hidden Markov Models, Recurrent Neural Networks, and Long Short Term Memories) and Hidden Markov Models (HMMs), whose revolutionary effects in CS subfields mirror the transformative influence of LLMs today and juxtapose their wide impact with that of LLMs. We compile all representative papers for RNNs-LSTMs and HMMs from [58], highly cited not just within NLP but also across other CS sub-fields, and count the citations they receive from non-CS fields using a methodology the same as Q1. To assess the diversity of their adoption, we use Gini indices on the citation distributions for these technologies. A Gini index of 0 indicates a perfectly equal distribution, whereas a value of 1 indicates total concentration in a single technology.
Results.
In Table 3, we present the diversity indices: HMMs (0.66), RNN/LSTMs (0.63), and LLMs (0.56). LLMs thus show the most even spread of adoption across fields.
Discussion.
LLMs demonstrate the lowest Gini index among these technologies, suggesting broader adoption across various non-CS fields than earlier CS exports. Further, to compare the popularity of LLMs with previous technologies, we set four different thresholds and show the number of fields surpassing each threshold. Fig 5 shows that LLMs receive citations more broadly across different fields, emphasizing their wider range of applications (refer to S3 Appendix) for additional analyses.
Number of fields surpassing the threshold number of citations for different NLP Technologies.
Evolving usage patterns of LLMs
To study how LLM utilization across various fields has evolved, we address two research questions: first, we identify the most widely used LLMs in these fields; second, we analyze which fields have embraced newer LLMs over time.
- (Q4) What are the most popular LLMs in different non-CS fields, and why?
Method.
We evaluate the popularity of LLMs across non-CS fields by computing the annual average number of citations each LLM receives from these fields.
Results.
Fig 6 shows the top 25 most popular LLMs across diverse non-CS fields (plot for a larger list of LLMs in S3 Appendix). Below, we summarize the main findings.
Discussion.
BERT, despite being officially published in 2019 (relatively early in the evolution of LLMs), remains the most popular LLM among non-CS fields, with the highest average citations per year. This could be because when BERT was introduced, there were fewer LLM options, leading to more citations. Consequently, many papers that incorporated LLMs during that period likely cited BERT, contributing to its higher average citation count. Secondly, the extensive, wide, and successful use of BERT likely made it a reliable choice for non-CS fields. In contrast, the behaviors and capabilities of newer LLMs are still undergoing exploration and evaluation.
In Fig 7, we show the number of LLMs published in each year. We observe that, in recent years, the proliferation of LLM research has led to the emergence of numerous new LLMs, and by February 2024, we already see citations beginning to shift from earlier dominant models toward these then-recent entrants (e.g., GPT-3, GPT-2, LLaMA, T5).
- (Q5) What is the average LLM citation age in non-CS fields, and how does it differ across fields?
Method.
The LLM citation age represents the average age of the LLMs that papers in a particular field are citing. A low citation age indicates a preference for newer LLMs, while a high age suggests reliance on older ones. Fields falling in between use a mix of both.
If a paper xi from research field c (i.e., ) cites an LLM paper yj, then the age of LLM citation (AoC) is taken to be the difference between the year of publication (YoP) of xi and yj:
We calculate the mean Age of LLM Citation (mAoCc) for a field c as:
Results.
Fig 8 shows that Psychology and Linguistics cite relatively recent LLMs (low mAoC), while Environmental Science relies more on older models. Biology and Chemistry strike a balance between older and newer models.
Discussion. These findings suggest that some fields (e.g., Psychology) rapidly integrate newer architectures, while others (e.g., Environmental Science) continue to depend on established models.
Psychology and Linguistics, having a longer history with LLMs, are now exploring newer LLMs, leading to low mAoC. Environmental Science relies mainly on older LLMs, possibly due to established performance and a lack of LLMs specifically created for the field. However, fields like Biology and Chemistry keep a balance between older LLMs while embracing newer LLMs, which are suited for their task, leading to moderate mAoC.
Applications of LLMs in non-CS fields
To examine LLM applications in non-CS fields, we qualitatively identify tasks where LLMs are applied, and additionally, using keyword searches, we roughly gauge the frequency with which the papers that cite LLMs in these fields mention ethical risks.
- (Q6) Do non-CS fields primarily fine-tune LLMs or use them in inference/zero-shot setting?
Method.
To estimate how LLMs are used in practice, we searched citing papers for mentions of fine-tune, zero-shot, and inference within abstracts. This filtering yielded 3,675 papers mentioning one of these terms. We then conducted manual validation to estimate the reliability of these signals. Specifically, we randomly sampled 50 papers mentioning fine-tuning and 100 papers mentioning inference/zero-shot to assess whether the terms were used in the context of actually applying LLMs.
Results.
Among the 3,675 papers mentioning either fine-tuning or inference, 479 explicitly referenced fine-tuning and 3,196 mentioned inference or zero-shot usage. Manual validation showed that 43 of 50 sampled fine-tuning papers (precision) indeed described adapting LLMs for domain-specific tasks, while 80 of 100 sampled inference papers (precision
) genuinely reported applying LLMs without retraining. Overall, papers describing inference usage outnumbered those describing fine-tuning by a factor of approximately 6.6 to 1, indicating that inference-based adoption is by far the dominant practice in non-CS fields.
Discussion.
These findings suggest that non-CS fields overwhelmingly adopt LLMs as task-agnostic tools in inference or zero/few-shot settings rather than investing resources in fine-tuning. This aligns with the preference for models like GPT-3 or LLaMA, which can be readily applied across tasks without domain-specific retraining. The slightly lower precision in detecting inference cases reflects the broader and more varied use of the term “inference” across disciplines, but the manual validation confirms that the dominant pattern holds robustly at scale. Taken together, this analysis complements our earlier observation (Q5) that non-CS fields favor adaptable, off-the-shelf models rather than fine-tuned variants.
- (Q7) How are LLMs utilized in non-CS fields?
Method. Understanding these field-specific tasks, where LLMs find applications, can help the NLP community better understand, adapt, and align LLMs and research with the distinct requirements of these communities. We examine LLM usage in various non-CS fields using trigrams in paper titles, abstracts, and citation contexts. By citation contexts, we precisely mean sentences that explicitly mention LLMs by name or citation [59]. We further process them to identify trigrams indicative of tasks using regex-based heuristics, such as removing the stop-words and manually filtering out irrelevant trigrams. Unlike [17], we use trigrams over bigrams, as trigrams offer richer contextual information beneficial for task identification within each field.
Results.
In Table 4, we highlight the top three most frequent trigrams that serve as representative tasks in papers citing LLMs for each of the non-CS fields. This table offers a glimpse into the specific problems LLMs are employed to address across various fields. In Table 5, we present frequent bigrams from these research papers, offering insights into the topics that interest these fields.
Discussion.
Non-CS fields employ LLMs primarily to solve domain problems rather than to analyze the models themselves, suggesting that LLMs are being integrated as tools into disciplinary workflows (Refer to S3 Appendix for additional analysis).
- (Q8) How frequently do research papers in non-CS fields mention ethical risks?
Method.
Until the preceding sections, we studied the impact of LLMs on non-CS fields and their utilization trends. However, it is essential to acknowledge that LLMs come with inherent risks, such as bias and hallucination [60]. These concerns are particularly relevant when LLMs are applied to sensitive applications.
While we acknowledge that pinpointing research papers that address ethical concerns about LLMs from a large collection presents a significant challenge (and requires huge manual efforts), we argue that identifying the contexts in which these papers discuss risks and ethical concerns can provide a rough idea and preliminary insights into how deeply non-CS fields are engaged with the ethical implications of LLMs. Hence, we examine papers that cite LLMs, aiming to identify contexts where discussions of risks and ethical considerations closely coincide with mentions of LLMs within the same sentence. Additionally, we extend our scrutiny to the titles and abstracts of these papers, assessing whether these concerns are underscored as particularly significant. Specifically, we searched for keywords related to ethical risks (e.g., “bias,” “hallucination,” “ethics”) and validated the method through manual checks of 100 flagged and 200 unflagged papers.
Results.
In Fig 9, we present the percentages of papers in each field that mention the risks associated with LLMs in their titles, abstracts, or citation contexts at least once relative to the total number of papers citing LLMs in that field. This acts as an approximate indicator of the relative importance of ethical considerations within various non-CS fields. To assess the reliability of our ethical-risk detection method, we manually reviewed 100 randomly selected papers on ethical concerns related to LLMs. The analysis indicated high recall, with only 7 relevant papers raising such concerns without explicitly referencing LLMs, confirming reliability. Refer to S2 Appendix for a detailed description and manual evaluation of this analysis method.
% of papers per field (Y-Axis) discussing LLM risks.
Our analysis reveals that, on average, approximately 2.01% of non-CS papers citing LLMs mention the ethical risks of LLMs. This finding is concerning, as it suggests that many non-CS fields are actively using LLMs without recognizing the ethical risks that these models entail.
Discussion.
Despite growing adoption, explicit acknowledgment of ethical risks remains rare outside a few disciplines, pointing to a substantial awareness gap as LLMs are applied in sensitive domains such as Medicine and Psychology.
Discussion: Opportunities and challenges
Our study provides the first large-scale descriptive baseline of LLM diffusion across disciplines beyond Computer Science. Through this study, we explored the broad diffusion of LLMs across academic disciplines beyond CS, highlighting their influence on fields as varied as Linguistics, Engineering, Medicine, and Law. By analyzing citation data and usage patterns from over 148k papers across 22 non-CS domains, we learned several important lessons about the evolving nature of LLM usage across scholarly research.
Firstly, we observe that LLMs have gained significant traction across multiple non-CS fields, with Linguistics, Engineering, and Medicine being the primary domains of citation. The growing number of LLM citations over time reflects the increasing adoption of these models, particularly those capable of zero-shot and few-shot learning, which do not require fine-tuning for domain-specific tasks. This widespread adoption highlights the versatility of LLMs in addressing diverse challenges, from protein function prediction in Biology to legal case retrieval in Law.
However, despite the broad use of LLMs, our study uncovers disparities in how various fields engage with this technology. Fields that are traditionally closer to NLP, like Linguistics, tend to exhibit higher levels of adoption, whereas other fields, such as Sociology or Philosophy, show more cautious or less frequent usage. This variation points to the need for more tailored LLM models that can better meet the specific needs of these fields.
Moreover, while LLMs are becoming increasingly popular, they come with their own set of concerns, particularly ethical issues such as the potential for biased or misleading outputs (and even hallucinations). Our analysis reveals that discussions around these risks are notably absent in many non-CS papers that cite LLMs, which raises alarms about the lack of awareness (consequently, mitigation strategies) across various disciplines. Fields like Philosophy and Law are exceptions, where ethical considerations are more actively addressed, but the overall trend indicates an urgent need for better integration of ethical frameworks in LLM usage.
As we look toward the future, it is clear that LLMs will continue to shape interdisciplinary research. The increasing volume of LLM citations across disciplines suggests that these models will not only augment existing research methods but also facilitate entirely new lines of inquiry. However, for this potential to be fully realized, it is crucial that the broader academic community, especially those in non-CS fields, become more educated about both the capabilities and risks associated with LLMs. Furthermore, there is a pressing need for the development of domain-specific LLMs that are better suited to the unique challenges of fields such as Medicine, Psychology, and Environmental Science.
In conclusion, while the integration of LLMs into non-CS research is undeniably a promising development, it brings with it both exciting opportunities and significant challenges. Researchers and practitioners must engage in responsible AI practices, fostering interdisciplinary collaboration and a deeper understanding of the ethical implications of LLMs. Recommendations include further investigation into the ethical aspects of LLM usage, the creation of tailored models for specific disciplines, and promoting awareness about the limitations of these models in handling complex, domain-specific tasks. By addressing these concerns, we can ensure that LLMs contribute to the advancement of knowledge in a safe, effective, and ethical manner.
Key messages
Our findings highlight several important takeaways for understanding the diffusion of LLMs across non-CS fields:
- Broad diffusion with disciplinary variation. LLMs have diffused widely beyond Computer Science, with Linguistics, Engineering, and Medicine emerging as the most active adopters. Fields more closely aligned with CS tend to adopt faster, whereas others show slower or more selective uptake.
- Enduring influence of early models. BERT remains the most enduringly cited LLM, while task-agnostic models such as GPT-3 and LLaMA are increasingly favored for domain-specific applications due to their zero/few-shot capabilities.
- Predominantly applied to domain tasks. Non-CS fields typically use LLMs to solve their own domain problems—such as protein function prediction, legal judgment retrieval, or radiology report generation—rather than analyzing the models themselves.
- Gaps in ethical awareness. Only a small fraction of papers outside CS explicitly acknowledge ethical concerns such as bias, hallucination, or reproducibility. This highlights a pressing need for broader awareness and integration of responsible AI practices across disciplines.
- Opportunities for tailored models. The variation in adoption patterns underscores the potential of developing domain-adapted or instruction-tuned LLMs better suited to the needs of Medicine, Environmental Science, Psychology, and other fields.
Threats to validity
While our study provides systematic evidence of the diffusion of LLMs across academic fields, several threats to validity should be considered when interpreting the findings:
- Citation data biases. Citations serve as a rough proxy for influence but do not always indicate direct methodological adoption. However, citation practices also differ across disciplines, with some fields citing more heavily than others. Although we partly mitigate this by reporting field-normalized proportions (e.g., share of papers citing LLMs within a field), these metrics cannot fully control for disciplinary citation cultures.
- Model selection bias. Our curated set of 106 LLMs prioritizes widely cited foundation models with identifiable publications. More recent instruction-tuned, multi-modal, or domain-specific LLMs are underrepresented because of limited citation coverage and uneven indexing in bibliographic databases.
- Field classification accuracy. Field labels were assigned using the S2FOS3 classifier, which has a reported accuracy of ∼86%. Misclassifications are inevitable and may affect fine-grained field statistics. To reduce noise, our analyses emphasize aggregate trends, and robustness checks with author primary fields yielded consistent patterns. Nonetheless, results should be interpreted at a field level rather than as precise counts for individual papers.
- Global and linguistic disparities. Semantic Scholar primarily indexes English-language and Western venues, which likely under-represents adoption in non-English research communities and regions with limited coverage.
- Measuring interdisciplinarity. Our proxy for interdisciplinarity—co-authorship with CS-affiliated authors—offers only a first-order approximation. Some non-CS researchers may hold CS training or dual appointments, which could inflate estimates of collaboration. More robust alternatives, such as cross-field publication history or venue diversity, remain promising directions for future work.
- Heuristics methods. Tasks extraction and ethical risk detection rely on heuristics and keyword searches. While small-scale manual validations suggest high precision and recall, these methods remain approximations. Noise in automatic extraction is expected, but aggregate-level results are robust enough to indicate broad trends.
By acknowledging these threats, we aim to contextualize our results as robust at the aggregate level but not immune to the biases inherent in large-scale scientometric analyses.
Supporting information
S1 Appendix. Supplementary definitions and discussions.
https://doi.org/10.1371/journal.pone.0337127.s001
(PDF)
S2 Appendix. Human evaluation of LLM citing papers that discuss LLM risks.
https://doi.org/10.1371/journal.pone.0337127.s002
(PDF)
References
- 1. Jurgens D, Kumar S, Hoover R, McFarland D, Jurafsky D. Measuring the evolution of a scientific field through citation frames. TACL. 2018;6:391–406.
- 2. Fortunato S, Bergstrom CT, Börner K, Evans JA, Helbing D, Milojević S, et al. Science of science. Science. 2018;359(6379):eaao0185. pmid:29496846
- 3. Bornmann L, Haunschild R, Mutz R. Growth rates of modern science: a latent piecewise growth curve approach to model publication numbers from established and new literature databases. Humanit Soc Sci Commun. 2021;8(1).
- 4.
Wahle J, Ruas T, Abdalla M, Gipp B, Mohammad S. We are who we cite: bridges of influence between natural language processing and other academic fields. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. p. 12896–913. https://doi.org/10.18653/v1/2023.emnlp-main.797
- 5.
Wahle JP, Lima RT, Abdalla M, Gipp B, Mohammad SM. Citation amnesia: on the recency bias of NLP and other academic fields. In: Rambow O, Wanner L, Apidianaki M, Al-Khalifa H, Eugenio BD, Schockaert S, editors. Proceedings of the 31st International Conference on Computational Linguistics. Abu Dhabi, UAE: Association for Computational Linguistics; 2025. p. 1027–44. https://aclanthology.org/2025.coling-main.69/
- 6.
Li S, Han C, Yu P, Edwards C, Li M, Wang X, et al. Defining a new NLP playground. In: Findings of the Association for Computational Linguistics: EMNLP 2023 . 2023. p. 11932–51. https://doi.org/10.18653/v1/2023.findings-emnlp.799
- 7. Radford A, Wu J, Child R, Luan D, Amodei D, Sutskever I, et al. Language models are unsupervised multitask learners. In: OpenAI blog. vol. 1; 2019. p. 9. https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf
- 8.
Henderson P, Li X, Jurafsky D, Hashimoto T, Lemley MA, Liang P. Foundation models and fair use. Journal of Machine Learning Research. 2023. p. 1–79.
- 9.
Wang Z, Li M, Xu R, Zhou L, Lei J, Lin X, et al. Language models with image descriptors are strong few-shot video-language learners. In: Advances in Neural Information Processing Systems. 2022. p. 8483–97. https://proceedings.neurips.cc/paper_files/paper/2022/file/381ceeae4a1feb1abc59c773f7e61839-Paper-Conference.pdf
- 10.
Wang X, Li S, Ji H. Code4Struct: code generation for few-shot event structure prediction. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023. https://doi.org/10.18653/v1/2023.acl-long.202
- 11. Movva R, Balachandar S, Peng K, Agostini G, Garg N, Pierson E. Large language models shape and are shaped by society: a survey of arXiv publication patterns. arXiv preprint 2023.
- 12.
Zhu S, Rzeszotarski J. What we talk about when we talk about LMs: implicit paradigm shifts and the ship of language models. In: Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025. p. 4628–46. https://aclanthology.org/2025.naacl-long.237/
- 13. Zhao WX, Zhou K, Li J, Tang T, Wang X, Hou Y. A survey of large language models. arXiv preprint 2023. https://arxiv.org/abs/2303.18223
- 14.
Singh J, Rungta M, Yang D, Mohammad S. Forgotten knowledge: examining the citational amnesia in NLP. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Toronto, Canada: Association for Computational Linguistics; 2023. p. 6192–208. https://aclanthology.org/2023.acl-long.341
- 15. Zhu X, Turney P, Lemire D, Vellino A. Measuring academic influence: not all citations are equal. Asso for Info Science Tech. 2014;66(2):408–27.
- 16.
Siddharthan A, Teufel S. Whose idea was this and why does it matter? Attributing scientific work to citations. In: Human Language Technologies 2007 : The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Conference. 2007. p. 316–23. https://aclanthology.org/N07-1040/
- 17.
Mohammad SM. Examining citations of natural language processing literature. In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020. p. 5199–209. https://doi.org/10.18653/v1/2020.acl-main.464
- 18.
Wahle J, Ruas T, Abdalla M, Gipp B, Mohammad S. We are who we cite: bridges of influence between natural language processing and other academic fields. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. p. 12896–913. https://doi.org/10.18653/v1/2023.emnlp-main.797
- 19. Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, et al. Attention is all you need. Advances in neural information processing systems. 2017;30.
- 20.
Devlin J, Chang MW, Lee K, Toutanova K. BERT: pre-training of deep bidirectional transformers for language understanding. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). 2019. p. 4171–86. https://aclanthology.org/N19-1423
- 21.
Raffel C, Shazeer N, Roberts A, Lee K, Narang S, Matena M, et al. Exploring the limits of transfer learning with a unified text-to-text transformer. In: The Journal of Machine Learning Research. 2020. p. 5485–551. https://jmlr.org/papers/volume21/20-074/20-074.pdf
- 22.
Brown T, Mann B, Ryder N, Subbiah M, Kaplan JD, Dhariwal P, et al. Language models are few-shot learners. In: Advances in neural information processing systems. 2020. p. 1877–901. https://proceedings.neurips.cc/paper/2020/file/1457c0d6bfcb4967418bfb8ac142f64a-Paper.pdf
- 23. Chowdhery A, Narang S, Devlin J, Bosma M, Mishra G, Roberts A, et al. PaLM: Scaling language modeling with pathways. J Mach Learn Res. 2023;24:1–113.
- 24.
OpenAI. 2023. https://openai.com
- 25. Touvron H, Lavril T, Izacard G, Martinet X, Lachaux MA, Lacroix T. LLaMA: open and efficient foundation language models. arXiv preprint 2023.
- 26.
Lo K, Wang LL, Neumann M, Kinney R, Weld D. S2ORC: the semantic scholar open research corpus. In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Online: Association for Computational Linguistics; 2020. p. 4969–83. https://www.aclweb.org/anthology/2020.acl-main.447
- 27.
Shi F, Chen X, Misra K, Scales N, Dohan D, Chi EH, et al. Large language models can be easily distracted by irrelevant context. In: International Conference on Machine Learning. PMLR; 2023. p. 31210–27. https://openreview.net/pdf?id=JSZmoN03Op
- 28. de Wynter A, Wang X, Sokolov A, Gu Q, Chen S-Q. An evaluation on large language model outputs: discourse and memorization. Natural Language Processing Journal. 2023;4:100024.
- 29.
Hall D, Jurafsky D, Manning CD. Studying the history of ideas using topic models. In: Proceedings of the 2008 conference on empirical methods in natural language processing. 2008. p. 363–71. https://aclanthology.org/D08-1038
- 30. Small H. Tracking and predicting growth areas in science. Scientometrics. 2006;68(3):595–610. pmid:32214554
- 31. Shibata N, Kajikawa Y, Takeda Y, Matsushima K. Detecting emerging research fronts based on topological measures in citation networks of scientific publications. Technovation. 2008;28(11):758–75.
- 32. Boyack KW, Klavans R. An improved practical approach to forecasting exceptional growth in research. Quantitative Science Studies. 2022;3(3):672–93.
- 33.
Prabhakaran V, Hamilton WL, McFarland D, Jurafsky D. Predicting the rise and fall of scientific topics from trends in their rhetorical framing. In: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Berlin, Germany: Association for Computational Linguistics; 2016. p. 1170–80. https://aclanthology.org/P16-1111
- 34. Grudin J. AI and HCI: two fields divided by a common focus. AI Magazine. 2009;30(4):48–57.
- 35. Liu S, Chen Y, Wei H, Yang J, Zhou K, Drucker SM. Exploring topical lead-lag across corpora. IEEE Trans Knowl Data Eng. 2015;27(1):115–29.
- 36.
Pramanick A, Hou Y, Mohammad S, Gurevych I. A diachronic analysis of paradigm shifts in NLP research: when, how and why?. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. p. 2312–26. https://doi.org/10.18653/v1/2023.emnlp-main.142
- 37.
Koch B, Denton E, Hanna A, Foster JG. Reduced, reused and recycled: the life of a dataset in machine learning research. In: Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2); 2021. https://openreview.net/forum?id=zNQBIBKJRkd
- 38. Mingers J, Leydesdorff L. A review of theory and practice in scientometrics. European Journal of Operational Research. 2015;246(1):1–19.
- 39. Chen C, Song M. Visualizing a field of research: a methodology of systematic scientometric reviews. PLoS One. 2019;14(10):e0223994. pmid:31671124
- 40.
Hansson B. Interdisciplinarity: for what purpose?. Policy Sciences. JSTOR; 1999. p. 339–43.
- 41. Leydesdorff L, Wagner CS, Bornmann L. Interdisciplinarity as diversity in citation patterns among journals: Rao-Stirling diversity, relative variety, and the Gini coefficient. Journal of Informetrics. 2019;13(1):255–69.
- 42. Van Noorden R. Interdisciplinary research by the numbers. Nature. 2015;525(7569):306–7. pmid:26381967
- 43. Truc A, Santerre O, Gingras Y, Claveau F. The interdisciplinarity of economics. Cambridge Journal of Economics. 2023;47(6):1057–86.
- 44. Porter AL, Rafols I. Is science becoming more interdisciplinary? Measuring and mapping six research fields over time. Scientometrics. 2009;81(3):719–45.
- 45. Shanahan M. Talking about large language models. Commun ACM. 2024;67(2):68–79.
- 46.
Sanh V, Webson A, Raffel C, Bach S, Sutawika L, Alyafeai Z, et al. Multitask prompted training enables zero-shot task generalization. In: International Conference on Learning Representations; 2022. https://openreview.net/forum?id=9Vrb9D0WI4
- 47.
Wei J, Bosma M, Zhao V, Guu K, Yu AW, Lester B, et al. Finetuned language models are zero-shot learners. In: International Conference on Learning Representations; 2022. https://openreview.net/forum?id=gEZrGCozdqR
- 48. Ganguli D, Lovitt L, Kernion J, Askell A, Bai Y, Kadavath S. Red teaming language models to reduce harms: methods, scaling behaviors, and lessons learned. arXiv preprint 2022. https://arxiv.org/abs/2209.07858
- 49. Rae JW, Borgeaud S, Cai T, Millican K, Hoffmann J, Song F. Scaling language models: methods, analysis & insights from training gopher. arXiv preprint. 2021. https://arxiv.org/abs/2112.11446
- 50.
Du N, Huang Y, Dai AM, Tong S, Lepikhin D, Xu Y, et al. Glam: efficient scaling of language models with mixture-of-experts. In: International Conference on Machine Learning. PMLR; 2022. p. 5547–69. https://proceedings.mlr.press/v162/du22c.html
- 51. Bommasani R, Soylu D, Liao TI, Creel KA, Liang P. Ecosystem graphs: the social footprint of foundation models. arXiv preprint 2023. https://arxiv.org/abs/2303.15772
- 52.
Lo K, Wang LL, Neumann M, Kinney R, Weld D. S2ORC: the semantic scholar open research corpus. In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020. https://doi.org/10.18653/v1/2020.acl-main.447
- 53. Baum LE, Petrie T. Statistical inference for probabilistic functions of finite state Markov chains. Ann Math Statist. 1966;37(6):1554–63.
- 54. Elman JL. Finding structure in time. Cognitive Science. 1990;14(2):179–211.
- 55. Hochreiter S, Schmidhuber J. Long short-term memory. Neural Comput. 1997;9(8):1735–80. pmid:9377276
- 56.
Rungta M, Singh J, Mohammad SM, Yang D. Geographic citation gaps in NLP research. In: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022. p. 1371–83. https://doi.org/10.18653/v1/2022.emnlp-main.89
- 57. Lerman RI, Yitzhaki S. A note on the calculation and interpretation of the Gini index. Economics Letters. 1984;15(3–4):363–8.
- 58.
Jurafsky D. Speech & language processing. Pearson Education; 2023.
- 59. Anderson MH, Lemken RK. Citation context analysis as a method for conducting rigorous and impactful literature reviews. Organizational Research Methods. 2020;26(1):77–106.
- 60. Kaddour J, Harris J, Mozes M, Bradley H, Raileanu R, McHardy R. Challenges and applications of large language models. arXiv preprint 2023. https://arxiv.org/abs/2307.10169