Can LLMs Simulate Child Users' Behaviors? A Case Study of Trust and Conformity Behaviors

Large language models (LLMs) are increasingly used to simulate child users in safety evaluations of child-facing AI systems. Yet it remains unclear whether LLMs can accurately reproduce children’s behavior, particularly patterns of trust and conformity that shape how children navigate novel information when interacting with AI systems. Prior work suggests that LLMs can reproduce behavioral patterns observed in social science experiments; however, existing evidence has focused primarily on adults or on children’s performance on linguistic and cognitive tasks. To address this gap, we examined whether frontier LLMs, prompted to act as child participants, could reproduce trust calibration and conformity behaviors observed among children using established paradigms from developmental psychology. In the trust study, human children were more likely to fact-check information after exposure to an unreliable source than after exposure to a reliable source, whereas none of the models reproduced this behavioral shift. In the conformity study, models tended to interpret social-norm messages as instructions, resulting in greater conformity than was observed among children and a failure to capture age-related developmental patterns. GPT 5.6 Sol, when prompted with personas augmented with our proposed Little Six personality attributes, achieved the closest overall distributional match to human responses in both studies. These findings reveal substantial limitations in the use of current LLM-based child simulators as proxies for children in AI safety evaluations.