A recent study published in Science reports that artificial intelligence (AI) tools like ChatGPT are dramatically increasing the volume of scientific papers. However, there’s a downside: although AI-generated text reads fluently, it adds little scientific value. The growing flood of such content is complicating peer review, funding decisions, and research oversight, as it becomes increasingly difficult to distinguish meaningful scientific contributions from low-value material.

Since ChatGPT became widely available in late 2022, many researchers have reported using these new tools to accomplish more work. At the same time, journal editors have observed a surge in submissions that are well-written but lack substantial scientific merit.

To investigate how large language models (LLMs) affect scientific publishing, a team led by Yin Yian (Yin Yi’an, phonetic) at Cornell University collected over 2 million preprints posted between January 2018 and June 2024 on three major preprint platforms—arXiv, bioRxiv, and the Social Science Research Network (SSRN)—spanning physical sciences, life sciences, and social sciences. These papers had not yet undergone peer review. The researchers compared papers presumed to be human-written before 2023 with AI-generated text. Using this comparison, they built a model designed to flag papers likely written with LLM assistance. With this detector, they estimated which authors might have used LLMs and tracked changes in the number of papers these scientists submitted—and whether those papers were accepted by scientific journals—before and after adopting AI tools.

The results show that LLM use significantly boosts research productivity. On arXiv, scientists flagged as likely LLM users posted about one-third more papers than those who did not use AI; on bioRxiv and SSRN, the increase exceeded 50%. The effect was especially pronounced for non-native English speakers. For example, researchers affiliated with institutions in Asia who used LLMs published 43.0% to 89.3% more papers across preprint servers than those who did not adopt the tool. Yin predicts this could ultimately reshape the global landscape of research productivity, helping regions previously hindered by language barriers catch up.

The study also notes potential advantages of AI tools in literature search and citation construction—they outperform traditional search tools in surfacing newer papers and relevant books. “LLM users gain access to more diverse knowledge, which might inspire more creative ideas,” said Keigo Kusumegi, the paper’s first author.

Although LLMs help produce more manuscripts, they also make it harder to identify truly high-quality scientific research. Yin warns that the growing gap between writing quality and research quality could have serious consequences. Editors and reviewers may struggle to spot the most valuable submissions, while universities and funding agencies may find that publication counts alone no longer reflect scientific contribution.

The researchers emphasize that these findings are observational. Next, they hope to use controlled experiments to test causal relationships.

A recent study published in Science reports that artificial intelligence (AI) tools like ChatGPT are dramatically increasing the volume of scientific papers. However, there’s a downside: although AI-generated text reads fluently, it adds little scientific value. The growing flood of such content is complicating peer review, funding decisions, and research oversight, as it becomes increasingly difficult to distinguish meaningful scientific contributions from low-value material.

Since ChatGPT became widely available in late 2022, many researchers have reported using these new tools to accomplish more work. At the same time, journal editors have observed a surge in submissions that are well-written but lack substantial scientific merit.

To investigate how large language models (LLMs) affect scientific publishing, a team led by Yin Yian (Yin Yi’an, phonetic) at Cornell University collected over 2 million preprints posted between January 2018 and June 2024 on three major preprint platforms—arXiv, bioRxiv, and the Social Science Research Network (SSRN)—spanning physical sciences, life sciences, and social sciences. These papers had not yet undergone peer review. The researchers compared papers presumed to be human-written before 2023 with AI-generated text. Using this comparison, they built a model designed to flag papers likely written with LLM assistance. With this detector, they estimated which authors might have used LLMs and tracked changes in the number of papers these scientists submitted—and whether those papers were accepted by scientific journals—before and after adopting AI tools.

The results show that LLM use significantly boosts research productivity. On arXiv, scientists flagged as likely LLM users posted about one-third more papers than those who did not use AI; on bioRxiv and SSRN, the increase exceeded 50%. The effect was especially pronounced for non-native English speakers. For example, researchers affiliated with institutions in Asia who used LLMs published 43.0% to 89.3% more papers across preprint servers than those who did not adopt the tool. Yin predicts this could ultimately reshape the global landscape of research productivity, helping regions previously hindered by language barriers catch up.

The study also notes potential advantages of AI tools in literature search and citation construction—they outperform traditional search tools in surfacing newer papers and relevant books. “LLM users gain access to more diverse knowledge, which might inspire more creative ideas,” said Keigo Kusumegi, the paper’s first author.

Although LLMs help produce more manuscripts, they also make it harder to identify truly high-quality scientific research. Yin warns that the growing gap between writing quality and research quality could have serious consequences. Editors and reviewers may struggle to spot the most valuable submissions, while universities and funding agencies may find that publication counts alone no longer reflect scientific contribution.

The researchers emphasize that these findings are observational. Next, they hope to use controlled experiments to test causal relationships.