TLDR: A large-scale empirical study analyzed over 2.1 million preprints from 2016-2025 to understand how generative AI (GenAI) and large language models (LLMs) are impacting academic publishing. The research found that LLMs have accelerated submission and revision cycles, increased team sizes, and subtly enriched linguistic complexity in abstracts. Most notably, there’s been a significant surge in AI-related research topics. However, the impact is highly uneven, acting as a ‘selective catalyst’ that disproportionately benefits computationally intensive fields, thereby widening existing disciplinary divides and raising new challenges for authorship and governance.
The world of academic publishing has seen significant changes over the last decade, with preprint repositories becoming crucial for sharing research before formal journal publication. Now, generative artificial intelligence (GenAI) and large language models (LLMs) like ChatGPT are introducing another layer of transformation, potentially altering how research papers are written and circulated. While many have speculated about this impact, concrete evidence has been scarce.
A recent study, titled Does GenAI Rewrite How We Write? An Empirical Study on Two-Million Preprints, delves into this question with a large-scale analysis. Researchers Minfeng Qi, Zhongmin Cao, Qin Wang, Ningran Li, and Tianqing Zhu examined over 2.1 million preprints published between 2016 and 2025 across four major repositories: arXiv, bioRxiv, medRxiv, and SocArXiv. Their goal was to systematically investigate how LLMs are reshaping scientific publishing by looking at changes in submission volume, authorship patterns, writing style, and the focus of research topics.
Accelerated Publishing Cycles
The study found that LLMs have indeed accelerated the pace of scientific publishing. After the release of major LLMs like GPT-3 and especially ChatGPT, there was a noticeable increase in the volume of submissions and a significant shortening of revision cycles across all repositories. This suggests that these tools are making the writing and editing process more efficient, allowing researchers to disseminate their work more quickly.
Evolving Collaboration and Productivity
Collaboration patterns are also shifting. The average number of authors per paper has steadily increased across all platforms, indicating a move towards larger research teams. While individual author productivity (papers per year) remained varied, the overall trend points to LLM-enabled workflows contributing to faster dissemination. This means that while more people might be involved in a paper, the tools help manage the increased workload.
Subtle Shifts in Writing Style
When it comes to the language and style of academic writing, the study observed subtle but important changes. Readability scores, like the Flesch Reading Ease, remained relatively stable, meaning the overall ease of understanding hasn’t drastically changed. However, there was an increase in lexical density and the ratio of difficult words in abstracts. This implies that abstracts are becoming more information-rich and are using more technical vocabulary, possibly reflecting a denser, more formal style of writing, especially in certain fields.
Surge in AI-Related Research
One of the most striking findings is the dramatic increase in AI-related research topics. After ChatGPT’s release, the share of documents focusing on AI surged significantly, particularly on arXiv, where it rose from under 5% to nearly 20% by mid-2025. This indicates that LLMs are not just tools for writing, but are also becoming a central subject of inquiry, influencing the intellectual direction of scholarship across various domains, including life and social sciences.
Uneven Impact Across Disciplines
The impact of GenAI is far from uniform. The study highlights that AI adoption is highly uneven across different disciplines. Computationally intensive fields, such as computer science and quantitative biology, showed the steepest increases in AI-related research and accelerated publishing. In contrast, fields like physics, mathematics, and some clinical or social science domains exhibited more modest changes. This suggests that LLMs act more as “selective catalysts,” amplifying existing strengths in fields already aligned with computational practices, rather than universally disrupting all areas of research.
Also Read:
- Unmasking AI’s Blind Spot: Fabricated Research Papers Fool LLM Reviewers
- Mapping Scientific Trends: An LLM Approach to Engineering Research in PNAS
Implications for the Future of Publishing
These findings have profound implications. Authorship is evolving, with LLMs blurring the lines between intellectual conception and text production. Human contributions may shift towards higher-level tasks like defining research questions and interpreting results. This also raises concerns about widening inequalities, as fields with established computational infrastructures gain disproportionate benefits. The study emphasizes the need for new governance frameworks to ensure trust, fairness, and accountability in an AI-enabled research ecosystem, addressing issues like disclosure policies and defining legitimate AI assistance.
