A comparative study of gendered object reasoning in children and large language models

Children often assign gender to everyday objects using symbolic schemas shaped by media, language, and lived experience. As generative AI becomes embedded in educational and domestic contexts, users increasingly prompt large language models (LLMs) to produce “child-like” explanations, yet such outputs may function more as cultural representations than developmental analogues. This study compares gendered object classifications generated by two LLM platforms (ChatGPT 5.1 and Gemini 3) with qualitative data from multilingual pre-adolescents (N = 21). Using a descriptive comparative design, ten familiar objects (five animate, five inanimate) were presented to each model under two conditions: baseline prompting and simulated-child prompting (“pretend you are an 11-year-old”). Model classifications and brief explanations were analyzed alongside children’s original responses. Simulated-child prompting increased surface-level alignment with children’s dominant classifications for both systems. However, the models’ explanations relied heavily on generalized aesthetic stereotypes, amplified affect, and invented autobiographical cues, diverging from children’s ego-centric and context-bound reasoning grounded in relationships, material culture, and personal memory. Platform differences were also evident: ChatGPT outputs showed greater symbolic flexibility and prompt sensitivity, while Gemini maintained stronger representational constraints and more frequent neutrality. The findings suggest that “child-voice” prompting can circulate flattened tropes of gender and childhood in school-adjacent settings. Implications are discussed for AI literacy education, classroom guidance on role-play prompting, and the design of child-facing generative AI systems.