Generative AI in UK higher education: Exploring its impact on student motivation, self-efficacy, and happiness to inform assessment design

Abstract

This mixed-methods study (N=201 UK undergraduates) identified a “well-being paradox” in AI-mediated learning. Contrary to expected outcomes, it found that greater reliance on generative AI tools for academic purposes was associated with higher academic well-being, especially positive emotion (β≈+0.33) and accomplishment (β≈+0.29). It also found that higher reliance on AI predicts higher AI-specific self-efficacy (β≈+0.33). The paradox arises as this reliance is also associated with substantially lower unaided academic self-efficacy (β≈−0.32; g≈0.90). This suggests that, for many students, confidence is shifting away from traditional, independent academic competence towards an outsourced, AI tool-supported form of performance. A qualitative analysis of Business undergraduates (n=15) went some way towards explaining this paradox. These students described AI as a safety net that reduces cognitive load and improves the quality of their output, resulting in genuine stress-relief and satisfaction. However, AI reliance was also seen to encourage strategic shortcuts and rising amotivation (β≈+0.18). Notably, another unexpected outcome was that guilt, which is typically linked to perceived misconduct or unearned accomplishment, is declining as AI use becomes normalised (β≈−0.28). This suggests a rapid recalibration of student norms associated with academic integrity. These relationships, however, are not uniform and the effects are discipline-dependent: Business students show resilient motivational patterns, whereas Humanities and Social Science students show a significant decline in intrinsic motivation (β≈−0.28). Across AI reliance terciles, a mid-reliance sweet spot emerges. Defined by self-imposed guardrails and metacognitive monitoring these students preserve academic autonomy and self-efficaciousness while enjoying AI-driven productivity gains. Taken together, these findings support assessment redesign that shifts from reactive detection toward structured, pedagogically driven AI integration.