Do LLMs Read for Free? Auditing the Open Access Bias in AI-Mediated Science

Abstract

How large language models affect various social activities has been widely discussed; however, how large language models themselves influence the mechanisms of scientific knowledge production remains unclear. To investigate this question, this paper designs a quasi-experimental study based on a large-scale bibliometric dataset. Using a Difference-in-Differences model and analyzing citation records from 64,841 articles across 120 communication journals, we find that the emergence of large language models is associated with a widening of the visibility gap between Open Access (OA) and non-OA articles. Specifically, OA articles experienced a significant post-ChatGPT citation acceleration of approximately 0.84% per month, accumulating to roughly 10% over 12 months. Importantly, this estimate represents a conservative lower bound, as gradual LLM adoption and journal review lags attenuate the measured treatment intensity. The findings are consistent with the view that, while LLMs may promote knowledge dissemination, they also risk exacerbating inequality in academic knowledge production—though the precise mechanism behind this pattern remains an important direction for future research.