From tool to partner: generative AI usage patterns and research performance among doctoral students

The rapid spread of generative artificial intelligence (GenAI) is reshaping doctoral education and raising new questions about its implications for research practice and performance. Using large-scale data from the 2025 Chinese National Doctoral Graduates Survey, this study examines how doctoral students adopt GenAI and how different modes of engagement are associated with research performance. Latent class analysis identifies three usage modes: Textual Tool (51.4%), Task Assistant (43.8%), and Thought Partner (4.8%). Ordered logit results show that deeper modes of engagement are more likely among students with STEM backgrounds, interdisciplinary experience, intrinsic motivation, and lower satisfaction with supervisory guidance. Associations between GenAI use and research performance vary systematically by mode and discipline. Textual Tool use is mainly linked to gains in publication productivity, whereas the more comprehensive integration represented by Thought Partner is associated with improvements in dissertation quality. Disciplinary contexts further condition these patterns. In abstract and conceptually oriented fields such as the humanities and mathematics, GenAI use is largely concentrated on text-related functions and is associated with modest, primarily quantity-based outcomes. By contrast, in computational and experimental domains including economics, physics, computer science, and engineering, deeper engagement through Task Assistant and Thought Partner use is more consistently associated with higher research productivity and quality indicators. Overall, the findings highlight that the implications of GenAI for doctoral research depend on how it is integrated into disciplinary research practices, underscoring the need for discipline-sensitive approaches to GenAI governance in doctoral education.