Introduction
The rapid advancement of Large Language Models (LLMs) has sparked a transformation across the academic spectrum. While their origins are rooted in Natural Language Processing (NLP) and Computer Science (CS), their influence has increasingly "leaked" into diverse disciplines. This study provides a systematic empirical analysis of how LLMs are permeating 22 non-CS fields, identifying early adopters, usage patterns, and the critical "ethics gap" in current research.
Methodology
The researchers utilized a multi-stage pipeline to quantify inter-field influence:
- Citation Mapping: Tracking citations from non-CS papers to 106 foundational LLM works (e.g., BERT, GPT-3, LLaMA).
- Metric Analysis: Employing the Gini Index to measure adoption inequality and Age-of-Citation (AoC) to assess the "freshness" of models used.
- Content Mining: Using NLP techniques to extract specific tasks and manual annotation to identify discussions of ethical risks (e.g., bias, hallucinations).
- Data Source: Analysis of ~148k papers and ~273k citations from the Semantic Scholar (S2) database, spanning 2018 to early 2024.
Key Insights: The Drivers of Adoption
The study identifies Linguistics, Engineering, and Medicine as the "Big Three" adopters, collectively accounting for a significant portion of LLM-related research outside CS. Linguistics and Engineering alone represent approximately 45% of all non-CS LLM citations.
A crucial finding is the role of interdisciplinary collaboration. There is a moderate correlation (0.38) between the presence of CS authors and the likelihood of LLM adoption in a paper. This suggests that computer scientists often act as "vectors," carrying technological expertise into other domains. Furthermore, non-CS researchers show a strong preference for "off-the-shelf" models like BERT, prioritizing stability and ease of use over the complex fine-tuning common in CS.
Breadth of Influence and the Ethics Gap
The adoption of LLMs is becoming increasingly widespread. The Gini Index, a measure of inequality, has shown a downward trend over the last several years, indicating that LLM influence is spreading more evenly across academia compared to previous technologies like RNNs.
Despite this broad adoption, a significant "ethics blind spot" exists. Only about 2% of papers citing LLMs explicitly discuss inherent risks such as hallucinations or algorithmic bias. While the CS community is deeply engaged in safety debates, other fields often treat these models as "black box" tools, potentially overlooking critical limitations.
Conclusion
LLMs have moved beyond being a niche CS interest to become a foundational tool for global research. From protein structure prediction in Biology to judgment prediction in Law, the applications are vast. However, the lack of ethical scrutiny in non-CS fields remains a concern. Future efforts should focus on developing domain-specific LLMs and "Responsible AI" frameworks that empower non-technical researchers to use these powerful tools safely and effectively.
Summary updated