Do AI agents demonstrate robust moral preferences in their decisionmaking?
How aligned, or misaligned, are these preferences to human values, goals, or intended outcomes?
How steerable are these preferences?
When and how do these preferences influence AI agents’ decisions in everyday, personal interactions?
The above questions are ones we ought to study as personal AI agents are adopted en masse. We build on the Moral Machine experiment, from Awad et al. in 2018, which collected more than 40 million responses across 230 countries on an autonomous vehicle-rendering of the perennial Trolley Problem. We examine frontier AI models’ choices in moral dilemmas, compare their revealed preferences with aggregate human preferences, and test how explicit value instructions change their decisions. We then explore model preferences across a range of everyday moral quandaries, from taking a sick child to the doctor but missing work, to confronting a close friend with ethical concerns about their partner. Across both life-or-death scenarios and the mundane, models reveal consistent preferences that can also shift with instruction, inviting further work on personal tools for measuring alignment and steering agent behavior.