AI tripled paper output but narrowed research topics by 4.63%

Artificial intelligence can make scientists more productive while narrowing the questions science asks. That is the paradox highlighted by two recent studies discussed by Park Jae-hyuk, a professor at KDI School of Public Policy and Management.

The first study examined 41.3 million natural-science papers published between 1980 and 2025. Researchers used a fine-tuned language model to identify work that incorporated AI methods, then tracked the effects on individual careers and the wider scientific community.

Scientists who used AI published 3.02 times as many papers each year as those who did not. Their work also received 4.84 times as many citations, and they became independent principal investigators an average of 1.37 years sooner. Yet the range of topics covered by AI-assisted research was 4.63% narrower than in other research. Follow-on studies drawing on that work also declined by 22%.

The analysis attributed the contraction to researchers gravitating toward mainstream subjects with abundant data and established academic interest, rather than exploring less-developed areas. Individual productivity expanded sharply, but the breadth of knowledge pursued across science contracted.

The second study examined how the spread of large language models affected research funding at the U.S. National Science Foundation and National Institutes of Health. After ChatGPT became widely used in 2023, researchers increasingly used LLMs to write proposals. Proposals with heavier LLM use showed lower semantic originality compared with recently selected projects. Moving from the bottom quarter to the top quarter of LLM use reduced originality scores by about 4 to 5 percentage points.

Despite that decline, proposals with greater LLM use had funding odds about 4 percentage points higher in NIH data and produced 5% more published papers after receiving grants. Those additional papers tended to attract relatively few citations rather than becoming highly influential studies.

Park argues that national research programs such as South Korea’s K-Moonshot and the United States’ Genesis Mission need evaluation systems that reward innovation and potential, not simply paper counts or short-term results. South Korea’s project aims to address 12 national missions across eight fields, including advanced biotechnology, quantum science and space, by 2035. The U.S. initiative has pledged more than $5 billion in federal investment. Park says research policy must guard against the uniformity AI can bring and support scientific diversity.

“

Can AI guarantee scientific innovation?