LLMs Are Reshaping Scientific Writing With Denser Vocabulary and Simpler Sentences

A large-scale study of scientific writing patterns has uncovered a marked shift in how researchers express themselves, with large language models emerging as a key driver of change. The research examined over 21 million scientific abstracts published between 2020 and 2024 alongside historical journal records to identify emerging trends in academic language.

The 2024 Turning Point#

The analysis pinpointed 2024 as a decisive moment when scientific writing underwent a striking transformation. Two linguistic changes occurred simultaneously: vocabulary became noticeably more complex and specialized, while sentence structure grew simpler and more direct. This pattern emerged consistently across all academic disciplines and across journals of different prestige levels, suggesting a widespread shift rather than isolated pockets of change.

Who Is Most Affected#

The linguistic restructuring proved especially pronounced among researchers writing in English as a second language, particularly those whose native languages have different grammatical structures from English. This group showed the most dramatic shift toward the new writing style, hinting that language models may be helping non-native English speakers meet the demands of publishing in English-language journals.

What LLMs Actually Do#

Controlled experiments that used language models to "polish" existing scientific text confirmed the hypothesis: these tools consistently produce the exact same pattern—introducing more specialized vocabulary while streamlining syntax into shorter, less complex sentences. This mechanical preference appears to be reshaping how scientific knowledge gets expressed across the board.

The Hidden Costs and Benefits#

The findings highlight a paradox baked into this shift. On one hand, the changes may help diverse scholars—particularly those from non-English language backgrounds—participate more fully in global scientific publishing. On the other, the move toward denser, more specialized language risks widening the gap between scientific discourse and language accessible to the general public. The research suggests that as LLMs become embedded in the scientific writing infrastructure, they may simultaneously democratize authorship while narrowing who can easily understand scientific findings.

Preprint: