The AI turn in science: a reflexive thematic analysis of scientists’ commentaries and educational implications

Background Artificial intelligence (AI) is increasingly embedded in scientific research, reshaping how knowledge is produced, evaluated, and legitimised. While much existing research examines AI as a technical innovation or pedagogical tool, less attention has been paid to how scientists themselves articulate its role within science.Purpose This study examines how scientists discuss AI in Nature and Science commentaries and considers implications for the nature of science (NOS) and science education.Sample The dataset comprises AI-related commentaries published in Nature (N = 242) and Science (N = 226) between 2021 and 2024, covering scientific discourse both prior to and following the public release of ChatGPT in late 2022.Design and methods Data were analysed using reflexive thematic analysis (RTA), with the Family Resemblance Approach to NOS (FRA – NOS) as an analytical orientation. This study is among the first to combine RTA with FRA – NOS to identify cross-cutting patterns between epistemic, methodological, social, institutional, political, and value-laden aspects of science. Such combination is methodologically resonant with Wittgenstein’s notion of family resemblance, foregrounding overlapping patterns of meaning rather than discrete categories.Result AI is not framed as a discrete tool but as a cross-cutting condition that reconfigures epistemic authority, infrastructure, scientific labour, governance, and public trust.Conclusion The analysis points to the importance of approaching AI in science education not merely as an add-on topic or technical skill, but as an opportunity to engage learners with science as a socially embedded, institutionally governed, and value-laden practice. Concrete examples are provided through educational applications of FRA – NOS. Methodologically, the study illustrates how combining RTA and FRA supports the identification of cross-cutting patterns that may be less visible in specific aspects of NOS or reliability-driven analyses.