Evaluating domain-specific moral reasoning in AI using the Adolescent Intermediate Concept Measure

This study examined how 110 large language models (LLMs) simulate adolescent moral reasoning. Model responses were compared to a normative human adolescent sample (n = 94), divided by developer origin (Western-trend vs. Eastern-trend), and examined with an unbiased-prompting modification and two moral-exemplar interventions (narrative guidance vs. persona adoption). Total ICM scores for LLMs ranged from 0.80 to 1.00 (M = 0.93, SD = 0.04). LLMs performed significantly better (M = 0.93) than human adolescents (M = 0.37) on all Ad-ICM indices (p < 0.001; Cohen’s d = 2.04–2.63). The Western-trend and Eastern-trend models performed nearly identically (Total ICM M = 0.94 vs. 0.93). Unbiased prompting resulted in small but reliable score decreases on five of seven indices (d ≤ 0.29), while moral-exemplar interventions shifted four of seven indices in small but significant ways (η2 = 0.03–0.07). However, there was no significant difference in the total ICM score between all three conditions. These findings indicate that current LLMs simulate adolescent-relevant virtue reasoning with a consistency that substantially exceeds human adolescent performance, while remaining largely, though not entirely, resistant to prompt framing and moral-exemplar manipulation, with implications for the responsible use of LLMs in moral education and youth-facing AI applications.