Information ethics education and artificial intelligence ethical judgment among university students

Generative artificial intelligence (AI) has become an everyday study tool for university students, and how students judge its acceptable academic use is now an explicit target of educational policy. This cluster-randomized experiment tested whether an information ethics education module improves university students’ AI ethical judgment, whether improvements persist over four weeks, and whether effects depend on students’ daily generative AI use. Sixteen intact classes from three universities in Zhengzhou, China, were randomly assigned at the class level to a four-week information ethics education module (eight classes, n = 120) or to regular coursework (eight classes, n = 120). AI ethical judgment was measured before the module, immediately afterward, and four weeks later with a double-rated dilemma test covering six academic scenarios and three dimensions (ethical sensitivity, evaluative judgment, behavioral intention); inter-rater ICC(2,1) values were 0.897, 0.923, and 0.930 at the three waves. The design-consistent primary analysis, a multilevel model with random intercepts for students and classes, yielded a significant arm by wave interaction, χ 2 (2) = 60.94, p < 0.001: composite judgment rose in the intervention arm and did not change in the control arm. Cluster-robust and wild cluster bootstrap analyses supported the immediate effect (b = 0.354, cluster-robust SE = 0.065; bootstrap 95% CI [0.140, 0.566]) and the delayed effect (b = 0.225, SE = 0.061; bootstrap 95% CI [0.062, 0.386]). Four weeks later, 59.9% of the immediate gain was retained in the intervention arm and 63.4% of the between-arm effect, estimates consistent with partial retention given their wide intervals. The intervention effect was larger among students who used generative AI more frequently (interaction b = 0.206, cluster-robust p = 0.040). A short, curriculum-embedded information ethics module thus improved university students’ AI ethical judgment, with gains partially retained after four weeks and the largest benefits among frequent users.