Scientific production in the era of large language models [pdf] (gwern.net)

Recent analysis has revealed that the adoption of large language models (LLMs) significantly boosts scientific productivity across various disciplines. Researchers assessed large datasets from three major preprint repositories and found that LLMs increase manuscript submission rates by 36.2% in arXiv, 52.9% in bioRxiv, and 59.8% in the Social Science Research Network. This enhancement in output is particularly pronounced among non-native English speakers, highlighting how LLMs can reduce barriers in scholarly communication and potentially shift the landscape of scientific authorship towards those in non-English-speaking regions. The implications of LLM use extend beyond just increased productivity, affecting perceptions of research quality as well. Traditional indicators like writing complexity may no longer reliably signal scientific merit, with LLM-assisted texts displaying counterintuitive trends—where more complex language is linked to lower peer-reviewed publication outcomes. This raises critical questions about the evolving standards of scientific communication in an era increasingly influenced by AI-generated content, urging policymakers to rethink the frameworks governing scientific institutions and the evaluation of research merit.