The rapid integration of artificial intelligence (AI) into higher education has renewed concern that technology-related strain may erode students’ higher-order thinking. Drawing on a social–cognitive framework, this study tested whether AI self-efficacy and technostress predict critical thinking in 340 Peruvian undergraduates, who completed three validated scales (Yoon’s Critical Thinking Disposition Inventory; the brief General Self-Efficacy Scale for AI, GSE-6AI; and the RED-Technostress scale). Data were modeled with PLS-SEM, with significance from 5000 bootstrap resamples and an out-of-sample predictive assessment (PLSpredict, CVPAT). AI self-efficacy was a robust positive predictor of critical thinking (β = 0.42, p < 0.001), whereas technostress had no direct effect (β = −0.06, p = 0.50); AI self-efficacy was associated with higher technostress (β = 0.24, p < 0.001), but that path did not carry over to critical thinking (no mediation). A hypothesized moderation (AI self-efficacy mattering more at higher technostress) was not confirmed: it reached significance under the two-stage method but not under a stricter bootstrap, and the hypothesis was therefore not supported. The model explained a modest share of variance in critical thinking (R2 = 0.17) and was invariant across public and private universities. These findings reposition technostress as, at most, a distal correlate and identify confident, literate engagement with AI, rather than the absence of strain, as the more promising lever for cultivating critical thinking.
The rapid integration of artificial intelligence (AI) into higher education has renewed concern that technology-related strain may erode students’ higher-order thinking. Drawing on a social-cognitive framework, this study tested whether AI self-efficacy and technostress predict critical thinking in 340 Peruvian undergraduates, who completed three validated scales (Yoon’s Critical Thinking Disposition Inventory; the brief General Self-Efficacy Scale for AI, GSE-6AI; and the RED-Technostress scale). Data were modeled with PLS-SEM, with significance from 5,000 bootstrap resamples and an out-of-sample predictive assessment (PLSpredict, CVPAT). AI self-efficacy was a robust positive predictor of critical thinking (β = 0.42, p < .001), whereas technostress had no direct effect (β = −0.06, p = .50); AI self-efficacy was associated with higher technostress (β = 0.24, p < .001), but that path did not carry over to critical thinking (no mediation). A hypothesized moderation (AI self-efficacy mattering more at higher technostress) was not confirmed: it reached significance under the two-stage method but not under a stricter bootstrap, and is therefore treated as preliminary. The model explained a modest share of variance in critical thinking (R2 = .17) and was invariant across public and private universities. These findings reposition technostress as, at most, a distal correlate and identify confident, literate engagement with AI, rather than the absence of strain, as the more promising lever for cultivating critical thinking.
University students’ adoption of generative artificial intelligence has become nearly universal, yet institutional guidance has not kept pace, leaving responsible use dependent on students’ own self-regulation. Drawing on social cognitive theory, this study tested whether self-regulated learning mediates the effect of AI literacy on the responsible use of generative AI, and how technostress influences that process. A cross-sectional survey of 820 Peruvian undergraduates from public and private universities was analyzed using covariance-based structural equation modeling; a two-stage moderated mediation model with latent interactions was estimated using latent moderated structural equations. Measurement invariance held across gender and university type, and common method variance was negligible. AI literacy predicted self-regulated learning (β = .626) and responsible use (β = .243), and self-regulated learning predicted responsible use (β = .445). The indirect effect was significant, β = .278, 95% CI [.226, .337], accounting for 53.4% of the total effect. Technostress did not moderate either the first stage or the direct path, but it significantly attenuated the second stage (b₂ = −.211, p < .001): the conditional indirect effect fell from .344 to .111 across moderator levels and ceased to differ from zero above 1.40 SD. Technostress does not prevent students from acquiring self-regulation; it prevents them from putting it into practice.
University students’ adoption of generative artificial intelligence has become nearly universal, yet institutional guidance has not kept pace, leaving responsible use dependent on students’ own self-regulation. Drawing on social cognitive theory, this study examined whether self-regulated learning accounts for the association between AI literacy and self-reported responsible use of generative AI, and how technostress relates to that association. A cross-sectional survey of 820 Peruvian undergraduates from public and private universities was analyzed using covariance-based structural equation modeling; a two-stage moderated mediation model with latent interactions was estimated using latent moderated structural equations. Measurement invariance held across gender and university type, and the available diagnostics did not indicate a dominant common-method factor. AI literacy predicted self-regulated learning (β = 0.626) and responsible use (β = 0.243), and self-regulated learning predicted responsible use (β = 0.445). The indirect effect was significant, β = 0.278, 95% CI [0.226, 0.337], accounting for 53.4% of the total effect. Technostress showed no significant moderation of the first stage or the direct path, but the second-stage path was significantly weaker at higher levels of technostress (b2 = −0.211, p < 0.001): the conditional indirect effect was 0.344 at low and 0.111 at high technostress and was not statistically distinguishable from zero above the estimated region-of-significance boundary of +1.40 SD. Technostress was not associated with weaker self-regulation, but with a weaker association between self-regulation and its behavioural expression. Because the design is cross-sectional, all estimates are reported as associations.