AI algorithms at the wheel: evaluating LLM responses to visual moral dilemmas in autonomous vehicles

This study examined how large language models (LLMs) respond to unavoidable life-or-death scenarios involving autonomous vehicles (AVs) and whether their moral preferences align with human judgements. We evaluated how 44 LLM variants responded to visual Moral Machine scenarios and compared their responses with those of human participants (n = 135). LLM responses were collected across three conditions: baseline, unbiased prompt, and character-based intervention conditions. Compared with human respondents, LLMs scored lower on Law (d = 0.68; humans more often spared lawful pedestrians), Age (d = 0.49; humans more often spared the young), and Status (d = 0.43; humans more often spared high-status individuals), and higher on Quantity (d = −0.38; LLMs more often chose the option that saved more lives). “Unbiased” prompting did not significantly shift LLM responses on any of the nine dimensions. Character-based interventions did not substantially shift LLM behavior across the measured dimensions, with the only observed change occurring in Species preferences. LLM responses in AV moral dilemmas diverged from human judgements in some dimensions while converging in others, suggesting both similarities and differences in moral preferences across groups. These patterns should be interpreted with caution and should not be considered sufficient evidence for AV programming decisions.