Artificial intelligence is helping scientists write papers faster than ever before, but the surge in productivity is clogging the academic system with “well-written but scientifically weak” research, according to a major new study from Cornell University.
The research, published in the journal Science, confirms what many editors have long suspected: while Large Language Models (LLMs) like ChatGPT are leveling the playing field for non-native English speakers, they are also making it increasingly difficult to distinguish between genuine scientific breakthroughs and “AI slop”.
“There’s a big shift in our current ecosystem that warrants a very serious look,” said senior author Yian Yin, an assistant professor of information science at Cornell. “It is a very widespread pattern, across different fields of science.”
The researchers analysed more than 2 million unreviewed papers (preprints) posted between 2018 and 2024 across three major repositories: arXiv (physics/maths), bioRxiv (biology), and SSRN (social sciences). By training an AI model to detect AI-generated text, they tracked scientists’ outputs before and after they adopted tools such as ChatGPT.
Stark results
The results were stark. On the arXiv platform, scientists who appeared to use AI tools boosted their output by about one-third. On biology and social science platforms, productivity jumped by more than 50%.
The biggest beneficiaries were researchers who do not speak English as a first language. Scientists from Asian institutions, for example, increased their paper output by up to 89% after adopting AI tools. The authors predict that this could trigger a global shift in scientific influence, empowering regions previously constrained by language barriers.
However, this boom comes at a cost. The study found a “disconnect” between linguistic polish and scientific substance.
Complex and clear
Traditionally, complex and clear writing has been a reliable proxy for high-quality research. But the study found that while AI-written papers often scored high on “writing complexity” tests, they were significantly less likely to be accepted by peer-reviewed journals compared to human-written papers of similar linguistic complexity.
In short, AI is helping researchers write convincing-looking papers that lack scientific rigour. This creates a headache for journal editors, funders, and policymakers who can no longer rely on productivity metrics or writing style to judge a scientist’s worth.
“The overall increase in AI-written papers is making it harder for many people… to separate the valuable contributions from the AI slop,” the authors warned.
Despite the flood of mediocrity, the study did find some bright spots. The use of AI-powered search tools (like Bing Chat) helped scientists find newer and more diverse citations, potentially driving more “creative ideas” by breaking them out of their usual citation bubbles.
“Already now, the question is not, ‘have you used AI?’,” said Yin. “The question is, ‘how exactly have you used AI and whether it’s helpful or not’.”
