A comprehensive Cornell University study reveals that artificial intelligence tools have dramatically increased scientific paper production while simultaneously undermining research quality, creating what researchers call a surge in "scientific AI slop." The findings, published in the journal Science, analyzed over 2 million academic papers and found productivity gains of up to 60 percent accompanied by declining acceptance rates at peer-reviewed journals.
The research team examined preprint papers posted between January 2018 and June 2024 across three major platforms: arXiv, bioRxiv, and the Social Science Research Network. By comparing pre-2023 human-authored papers with AI-generated text, researchers developed an algorithm to detect likely AI usage in scientific writing and track how adoption affected individual researchers' output and publication success.
Productivity Surge Across All Disciplines
Scientists using large language models like ChatGPT experienced dramatic productivity increases across all fields studied. Social sciences and humanities saw the largest jump at 59.8 percent, while biology and life sciences increased 52.9 percent. Physical sciences experienced gains ranging from 36 to 50 percent depending on the specific discipline.
"There's a big shift in our current ecosystem that warrants a very serious look," said Yian Yin, assistant professor at Cornell's Ann S. Bowers College of Computing and Information Science and senior author of the study. The productivity boost proved especially pronounced for non-native English speakers, with Asian researchers showing increases ranging from 43 to 89.3 percent.
Complex Language Masks Weak Research
The study uncovered a troubling reversal in traditional quality indicators. For human-authored papers, complex language correlated with higher publication rates. However, AI-assisted papers showed the opposite—the more complex the language, the less likely the paper achieved publication.
"AI-generated complex language was used to hide the low quality of the scholarly work," researchers concluded. This challenges the academic practice of using writing quality as a proxy for research value.
Broader Literature Access Through AI Search
The study did identify one positive development. Researchers compared article downloads from Google versus Microsoft's AI-powered Bing Chat, introduced in February 2023. Bing users accessed a wider variety of sources and discovered more recent publications compared to traditional Google search.
"People using LLMs are connecting to more diverse knowledge, which might be driving more creative ideas," said Keigo Kusumegi, doctoral student and first author. This finding suggests AI-enhanced search could promote interdisciplinary research, though researchers plan future studies to confirm whether broader access translates to genuine innovation.
Implications for Scientific Publishing
The research team warns that the flood of polished but low-value submissions threatens scientific integrity. Co-author Paul Ginsparg, physics professor and founder of arXiv, noted the challenge facing the research community as AI becomes embedded in standard tools like word processors and email applications.
Researchers propose several responses, including implementing deeper peer review focused on research methods rather than writing polish. One suggestion involves "fighting fire with fire" by deploying AI-powered review tools, similar to systems recently published by Stanford's Andrew Ng, to help manage submission volumes and identify substantive contributions.
The study included contributions from Xinyu Yang, Mathijs de Vaan, and Toby Stuart of UC Berkeley, with support from the National Science Foundation. Cornell will host a symposium March 3-5, 2026, examining generative AI's impact on research and developing policy guidance for the rapidly evolving landscape.
