The Artificial Moral Machine: Do LLMs Mirror Human Moral Diversity or Merely Caricature It?

Abstract

Persona conditioning is the practice of prompting LLMs to adopt specific demographic identities. Although it is increasingly used to simulate human perspectives, its fidelity in the moral domain is untested at scale. We address this gap by testing moral fidelity on three LLMs of different sizes: gpt-4.1-nano (the smallest in OpenAI's GPT-4.1 family), DeepSeek V3.1 (671B-MoE), and Llama 3.3 70B. Each model answers the same 576 cohorts of 5,356 trolley-style dilemmas, with 20 prompt variants per dilemma on nano (61M+ judgments) and one on each larger model (~3M each). Each cohort combines 7 demographic features (age, gender, education, income, politics, religion, geography); each dilemma tests 5 moral dimensions (sparing the young, women, humans, the high-status, the lawful). We compare the answers against the Moral Machine human dataset. All three tested models show a political master switch: political ideology explains 7.3 to 8.0 times more variance in moral preferences than geography. This is in contrast to the human pattern observed in the Moral Machine, where cultural clusters drive more variation than political views. The LLMs match human demographic effects no better than chance across all three models. Grouping the geographic labels of our tests into the same three cultural clusters used in the human Moral Machine does not reproduce the differences Moral Machine reported. Moreover, the 5 moral dimensions collapse onto a single axis in nano (meaning that across cohorts the dimensions move together in a single block), while in the two larger models they remain more separable. We further test robustness with a no-persona baseline and by removing intensifier wording ("strongly", "deeply") from the persona descriptions. We release the full dataset and analysis framework for replication.