Generative AI-supported peace pedagogy in higher education: A two-wave quasi-experimental study of AI literacy, intercultural sensitivity, and empathy

Abstract

Generative AI is increasingly used in higher education, but its value for humanistic learning depends on pedagogical designs that preserve learner judgment and require critical evaluation of generated content. This two-wave quasi-experimental study examined a six-week GenAI-supported peace-pedagogy module in 522 Saudi higher-education students (intervention n = 276; waitlist n = 246). Intercultural sensitivity, empathy, and AI literacy were assessed at pretest and posttest using the ISS-15, Interpersonal Reactivity Index, and an adapted Arabic administration of the Meta AI Literacy Scale (MAILS), respectively. Primary analyses used linear mixed-effects models for Time × Group interactions, with baseline-adjusted posttest ANCOVA models as sensitivity analyses; network analyses were secondary and exploratory. The groups differed at baseline on all three primary outcomes. The intervention group showed a larger increase in intercultural sensitivity (interaction b = 0.699, 95% CI [0.524, 0.875], BH-adjusted p < .001), and the baseline-adjusted posttest difference remained substantial (b = 1.042, 95% CI [0.918, 1.165], BH-adjusted p < .001). AI literacy showed a similar, smaller pattern (interaction b = 0.333, 95% CI [0.148, 0.517], BH-adjusted p < .001; baseline-adjusted posttest b = 0.639, 95% CI [0.534, 0.744], BH-adjusted p < .001). Empathy did not show robust evidence of improvement: the Time × Group interaction was negative (b = −0.252, 95% CI [−0.409, −0.095], p = .002), whereas the baseline-adjusted posttest group difference was not significant (b = 0.034, 95% CI [−0.083, 0.151], p = .572). Exploratory network comparisons did not detect significant group differences at either wave. Overall, the intervention group showed larger gains in self-reported intercultural sensitivity and AI literacy, whereas empathy showed no robust positive pattern. Because allocation was nonrandom, groups differed at baseline, attrition could not be reconstructed, and the comparison group received usual coursework, the findings cannot establish a causal effect or isolate GenAI’s contribution.