Modelling Human Generative AI Interaction in Higher Education: Evaluation Among Ghanaian Students

Abstract

Background: Generative artificial intelligence (GenAI) is rapidly transforming higher education, yet the cognitive processes through which it affects learning remain under-theorised. Objective: This study extends adoption-focused perspectives by developing and empirically testing a Human-AI Interaction Model (HAIM), which positions the quality of cognitive interaction, rather than mere adoption, as the key determinant of educational outcomes. The model draws on Extended Cognition Theory and Self-Regulated Learning Theory and specifies ten hypothesised pathways among seven constructs: AI literacy, trust in AI, cognitive offloading, co-creation behaviour, verification behaviour, learning outcomes, and critical thinking. Methods: Cross-sectional survey data were collected from 623 university students across five Ghanaian institutions and analysed using Partial Least Squares Structural Equation Modelling (PLS-SEM) with 5,000 bootstrap resamples. The measurement model exhibited good reliability, and convergent and discriminant validity were confirmed. Results: AI literacy positively predicted co-creation behaviour (β = .479) and negatively predicted cognitive offloading (β = -.199). Uncalibrated trust was positively associated with cognitive offloading (β = .319), which in turn reduced learning outcomes (β = -.216). Co-creation behaviour was the strongest predictor of learning outcomes (β = .531), while verification behaviour was the strongest predictor of critical thinking (β = .540). Mediation analysis confirmed that cognitive offloading and verification behaviour are the primary mechanisms linking trust and literacy to outcomes. Conclusion: The findings support a cognitively informed account of human-AI interaction and underscore the need for AI literacy interventions that cultivate critical engagement rather than passive dependence, particularly within Global South higher education contexts.