STHC factors on technology adoption. Adoption is neither exogenous nor random in our data; it is a choice. We observe correlations between STHC on one hand and AI adoption on the other, and we interpret these patterns in light of the conceptual framework developed in Section 2.

3.3.1 Matching Technology available at a time t, age of the researcher, initial training and research trajectory influence the opportunity to use AI. Hence, when matching individuals with similar potential to use AI, we take into account technology advancement and its applicability in a given field. As shown in Figure 3, we construct two matched samples. The first matching is used to investigate first-use of AI. Each focal scientist is matched with a non-focal scientist who (i) never published an AI paper, but published her first paper in the same year and same scientific field 16 Figure 3: Matching procedures to investigate first-use of AI (top) and re-use of AI (bottom) as the focal scientist, (ii) published a non-AI paper in the same scientific field and same year as the focal scientists first AI paper, and (iii) published at least one paper subsequently, as the focal scientist. Since our focus is on domain scientists, individuals with computer science papers are

The second matching is used to investigate re-use of AI. Here we compare two focal scientists who first published a paper on AI in the same scientific field and in the same year, but one of them reuses AI in a later year and the other does not. As for the first matching, we also require that both individuals published their first paper in the same field and year. We performed exact matching using all Openalex data, matching 23,918 first-users of AI with non-users of AI and connecting 13,211 re-users with first-users that did not re-use AI.

4 Empirical analysis This section provides at first some aggregate statistics on the diffusion of AI in research and discusses the main patterns of its adoption process. Detailed descriptive statistics are available in Appendix A. The section then focuses on the main econometric results regarding the relationship between STHC endowment, first-use of AI and its subsequent re-use. The analysis is broken down by scientific field in Appendix B. 4.1 Main trends on AI adoption in science A growing number of players: The adoption of AI technology by domain scientists is on the rise, with over 20,000 individuals incorporating it for the first time in 2020 alone. This pattern aligns 17 Table 1: Co-authors of first-time AI adopters non-AI papers AI papers t-test # Authors # CS aut. # AI exp. aut. # Domain aut. # Newbies aut. 10.73 (10.31) 2.19 (4.09) 1.71 (3.48) 8.53 (7.65) 1.19 (1.70) 11.65 (12.51) 3.12 (4.82) 2.84 (4.54) 8.53 (9.62) 1.41 (2.82) 14.32 36.66 49.34 0.04 16.13