Evaluating Empathetic Decision-Making in AI: A Comparative Study of Open-Source Models in High-Stakes Scenarios

Artificial intelligence is increasingly used in high-stakes situations such as military operations, medical triage, and humanitarian crises. While these systems optimise accuracy and efficiency, they often fail to account for empathy, which is essential for ethical decision-making. This study compares the empathetic reasoning abilities of open-source AI models, including GPT-J, LLaMA 2, BLOOM, decision tree, and random forest classifiers, across curated military, medical, emotional, and India-specific moral dilemmas. Models were evaluated on decision accuracy, empathy alignment score, explanation quality, and consistency. ANOVA results show no significant difference in decision accuracy (F=0.55, p=0.703) but highly significant differences in empathy alignment (F=121.99, p<0.001) and explanation quality (F=26.92, p<0.001). GPT-J demonstrated the highest empathy alignment, while interpretable models showed lower empathy but higher consistency. The findings highlight the need for hybrid systems that combine empathetic reasoning with stable and interpretable decision pathways.