Using generative AI in science necessitates that institutions and researchers practice epistemic reflexivity to protect intellectual diversity and resist the threat of scientific monoculture. We argue that researchers and institutions can position generative AI both as a tool for epistemic reflexivity and as an object that must be examined reflexively. Traberg et al.1 cautioned the scholarly community that using generative AI in research threatens to turn it into a “scientific monoculture” by nudging researchers toward an algorithmic average, a gravitational pull toward what has already been done.
This risk is disproportionate for multiple groups of researchers. Early-career researchers who have not yet developed robust expertise are more vulnerable than later-career researchers. Researchers from the Global South might find that their work is filtered through the Western-centric ways of thought embedded within the foundational training corpora of most LLMs.
Researchers making original contributions may find their work penalized simply because it departs from the patterns in an LLM’s training data, disadvantaging precisely the most original contributions. This systemic exclusion of minority paradigms acts as a mechanism of epistemic narrowing. Furthermore, such threats may compound over time.
For example, an uptake of generative AI use in research might lead to convergent research products re-entering training corpora, causing a feedback loop that will tighten even more across successive model generations. Currently, institutional guidance - whether that comes from universities, professional associations, or publishers - frames training and compliance in AI-assisted research as an individual obligation, primarily focusing on maximizing individual benefits and limiting the potential of negative repercussions to the individual2,3. However, science homogenization is not something any individual researcher can prevent through their own choices and actions alone.
Rather, addressing homogenization requires both individual practices to protect oneself from limiting applications of generativeAI and institutional policies to encourage divergence. The time to intervene is now, before current institutional guidance becomes normalized and individual behaviors become disciplinary habits. We advance that preventing scientific monoculture requires positioning generative AIboth as a tool for epistemic reflexivity and an object that must be examined reflexively at individual and institutional levels.
Epistemic reflexivity is the practice of interrogating not just what we know but how we know it, namely our disciplinary training, theoretical lenses, and methodological paradigms that shape which questions we ask, what methods we choose, and what counts as evidence4,5. Beyond optimizing for individual skills and efficiencies6, we propose the following individual strategies for epistemically reflexive GenAI usage: Consider whether generative AI is used in aide mode, where the impact of the AI’s subjectivity is limited (e.g., when editing language), or in partner mode, where AI’s subjective perspective heavily shapes the user’s thought and production (i.e., when generating drafts, ideating, interrogating, analyzing, or generating arguments and responses). Adopting an epistemic reflexivity stance empowers researchers to distinguish these modes and operate cautiously when epistemic risk is high.
Extract — continue reading at the source.
