AI in science: When and where it makes a difference

Why do some scientific fields benefit far more from AI than others? To answer this question, we study the diffusion and effects of AI across more than 170 scientific fields between 2005 and 2023, drawing on a dataset of over 80 million scientific publications, and assessing its contribution to knowledge creation using indicators of novelty and impact. We find that, on average, the use of AI is associated with more novel and highly cited research, and that these effects are strengthened with recent technological advances (namely, transformers and large language models). However, these benefits are far from uniform. The contribution of AI, especially in terms of novelty, intensifies with its penetration within a field, that is, the extent to which AI is widely adopted and embedded in research practices, and with the “roughness” of the underlying knowledge space, meaning the degree to which ideas are fragmented and combinatorially complex. • AI positively contributes to scientific novelty and impact. • Effects strengthened with recent AI advances. • Effects are uneven across 170+ fields. • The contribution of AI intensifies with its penetration within a field. • Strongest gains occur in fragmented (“rough”) knowledge spaces.