AI HAS CHANGED SOME WORDS OF SCIENCE, NOT SCIENCE ITSELF — YET
Since large language models became publicly available at the end of 2022, researchers have used them to write, edit and search the literature. Studies credit them with real gains for individual scientists, in productivity and visibility. But worries have grown about science as a whole: an “AI paradox” in which productivity comes at the expense of quality, a flood of “AI slop”, and a creeping monoculture that would narrow the range of ideas.
Ron Bekkerman, Nichole Li, Bahareh Kargar and Mark Somers, at the New Jersey Institute of Technology, set out to measure whether anything had broken — not against an idealized past, but against the trends that were already under way.
Ten years of full texts
Their corpus holds about 2.5 million full-text papers published between 2015 and 2025, in five disciplines: computer science, mathematics and physics (from arXiv), and health and life sciences (from PubMed Central). Rather than looking at abstracts only, they analysed whole papers, along a ladder of questions: which words are used, how rich and technical the language is, how diverse the topics are, how papers are built, and who writes them.
The method is the same throughout: compare each trend before and after 2022, and look for a break.
Words that rose and fell
Starting from the 20,000 most common English words, reduced to 14,484 base forms, the team found 796 that surged after 2022, 199 that declined and 71 that rose and then dropped. The risers are polished analytical words that language models favour, such as “establish” and “underscore”; the losers are plain ones such as “good” and “happen”. The famous “delve” shows the rise-and-drop pattern: nearly absent before 2023, it peaked in 2024 and fell sharply in 2025 — as models were updated and readers grew wary, the authors suggest.
The signal is strongest in computer science and weakest in mathematics, whose writing is held in check by formal notation. And it is not confined to abstracts: the surging words appear in the body of papers too.
No dilution of scientific language
Yet the words concerned — a little over a thousand — are a drop in a vocabulary of tens of millions of distinct words. Lexical diversity, measured over sliding windows of 50 words, shows no significant change of slope in 2022 in any discipline. And the density of scientific terminology keeps rising in all five fields. In the authors’ words, “scientific papers appear more scientific over time”.
Topics: two false alarms
Do papers increasingly talk about the same things? In physics and mathematics, topic diversity is flat. In computer science, it falls almost in a straight line — but the decline began long before 2022, as the field converged on deep learning and generative AI, and it has even slowed slightly since 2021.
Health and life sciences did show a sharper decline after 2022. But much of the similarity between papers came from their format: standard blocks on licences, disclosures and contributor roles, as biomedical literature moved towards a unified format. Filtering out 105 such boilerplate phrases removed most of the decline.
Papers built to a template
The structure of computer science papers, studied on nearly 438,000 of them, has changed a lot — but steadily. Papers with only one to three sections fell from 17% to 4%, while those with eight or more rose from 5% to 30%. Introductions shrank from 2,141 to 1,236 words on average, while reference lists nearly doubled in length. None of these trends broke in 2022.
Bigger teams, more papers — mostly as before
Teams of co-authors had been growing in every discipline before 2022, and simply kept growing. One exception stands out: computer science teams from Asia grew faster after 2022. The number of papers also grew faster in computer science and physics after 2022, which the authors associate with language models. A jump in biomedical output in 2023 is more likely a rebound from a slump in 2022 caused by COVID-era disruption.
Not yet
The authors find little evidence of the “AI paradox” at the scale of science as a whole. They warn against reading a handful of fashionable words as proof of a revolution, and argue that policy — on disclosure, on peer review — is the right tool to manage the tension. Their reassurance has a time limit, though. The study covers only three years after 2022 and measures the observable features of writing, and the authors find “the speed with which LLMs are penetrating the research process” concerning.
