Generative AI use profiles and academic outcomes in higher education: the role of institutional support

Generative AI (GenAI) is increasingly embedded in higher education as an interactive learning environment (ILE), yet evidence on learning benefits remains mixed. Using an anonymous survey of 2,240 undergraduates from a public university in central China, we adopt a person-centred approach to test whether latent GenAI engagement profiles predict broader perceived capability development beyond overall usage intensity, and whether institutional support breadth relates to broader gains across profiles. Capability development is measured by AbilitySum (1–4), counting self-reported improvements in critical thinking, creativity, teamwork, and self-management. Latent class analysis identifies five profiles differing in purpose coverage, tool ecosystems, and reliance. Ordered logit models show profile membership predicts AbilitySum net of usage intensity. Profile-stratified models indicate SupportCount (0–4 modalities) is positively associated with AbilitySum in every profile (AMEs ≈ 0.16–0.29), with the largest marginal returns for mainstream regular learners; bootstrap contrasts reveal a significant Class 3–Class 5 difference after Holm adjustment. An XGBoost robustness check corroborates a monotonic support–gain pattern. Conceptually, use architecture clarifies how student interactions with generative AI are linked to competence development through the learning process rather than simply usage intensity. Thus, it expands on self-regulated learning (SRL) theory by highlighting how purpose, tool, and dependency configurations shape learning. Findings highlight the importance of “how” GenAI is engaged and suggest universal-plus-tiered institutional scaffolding to translate GenAI access into broad capability development.