Latent Classes of Generative AI Use and Mental Health Among U.S. College Students

Background and Purpose: Generative AI is becoming embedded in college students’ academic and everyday lives, yet students do not use these tools in uniform ways. Existing discussions have focused largely on academic integrity, learning, or overall adoption, but less is known about whether broader, multi-domain patterns of use are associated with student well-being. Because students use generative AI across academic, social, health-related, and recreational contexts, person-centered approaches may clarify whether distinct patterns of engagement are linked to mental health outcomes (Masyn, 2013; Nylund-Gibson & Choi, 2018). This study identified latent profiles of generative AI use among U.S. college students and examined whether profile membership was associated with depressive symptoms, anxiety symptoms, and loneliness.

Methods: Data came from the 2024–2025 Healthy Minds Study, a web-based survey of students at participating U.S. colleges and universities. Analyses were limited to respondents who completed the elective module on Knowledge and Attitudes about Artificial Intelligence (N=953). Latent class analysis identified a five-class solution based on eight indicators of generative AI use: academic use when allowed, academic use when forbidden, non-academic work, messaging friends and family, physical health, mental health, companionship, and creating content for fun. The five classes were labeled Minimal, Light, Moderate, Regular, and Heavy users. Mental health outcomes included depressive symptoms (PHQ-9), anxiety symptoms (GAD-7), and loneliness (UCLA 3-item scale). Posterior-probability-weighted linear regression models examined associations between class membership and each outcome, with Light users as the reference group. Adjusted models controlled for sociodemographic, academic, environmental, AI-related, and stigma-related factors.

Results: The five-class solution differentiated students by breadth and intensity of generative AI use. Compared with Light users, Moderate users had significantly higher depressive symptoms (b=2.06, 95% CI [1.49, 2.64], p<.001), anxiety symptoms (b=1.49, 95% CI [0.91, 2.07], p<.001), and loneliness (b=0.75, 95% CI [0.56, 0.95], p<.001). Heavy users also reported significantly higher depressive symptoms (b=2.00, 95% CI [1.20, 2.79], p<.001), anxiety symptoms (b=2.74, 95% CI [1.94, 3.54], p<.001), and loneliness (b=0.40, 95% CI [0.14, 0.67], p=.003) than Light users. Regular users reported slightly higher anxiety (b=0.48, 95% CI [0.02, 0.95], p=.042) but did not differ on depressive symptoms or loneliness. Minimal users did not differ significantly from Light users on any outcome. Overall, Moderate and Heavy users showed the most consistent elevations across mental health outcomes.

Conclusions and Implications: Associations between generative AI use and student well-being may depend less on whether students use AI at all than on how broadly and intensively they use it across domains. Students in the Moderate and Heavy profiles reported the greatest mental health burden, whereas the Light profile, characterized primarily by allowed academic use, showed comparatively more favorable outcomes. For social work education and higher education practice, these findings underscore the importance of moving beyond user versus nonuser distinctions and considering how distinct patterns of AI engagement relate to student well-being. As institutions develop AI guidance, policies focused only on academic integrity may miss broader student support concerns, including how students use AI in coping, information-seeking, and everyday functioning.