Cross-Cultural Safety Judgments in Child Environments: A Semantic Comparison of Vision-Language Models and Humans

Despite advances in complex reasoning, Vision-Language Models (VLMs) remain inadequately benchmarked for safety-critical applications like childcare. To address this gap, we conduct a multilingual (English, French, Polish, Japanese) comparison of VLMs and human safety assessments using a dataset of original images from child environments in Japan and Poland. Our proposed methodology utilizes semantic clustering to normalize and compare hazard identification and mitigation strategies. While both models and humans identify overt dangers with high semantic agreement (e.g., 0.997 similarity for ‘scissors’), their proposed actions diverge significantly. Humans strongly favor direct physical intervention (‘remove object’: 64.% for Polish vs. 55.0% for VLMs) and context-specific actions (‘move object elsewhere’: 17.8% for Japanese), strategies that models under-represent. Conversely, VLMs consistently over-recommend supervisory actions (such as ‘Supervise children closely’ or ‘Supervise use of scissors’). These quantified discrepancies highlight the critical need to integrate nuanced, human-like contextual judgment for the safe deployment of AI systems.