Large-scale moral machine experiment on large language models

TL;DR

Evaluating moral judgments across 52 different LLMs, including multiple versions of proprietary models and open-source alternatives and open-source alternatives, shows that proprietary models and open-source models exceeding 10 billion parameters demonstrated relatively close alignment with human judgments, with a significant negative correlation between model size and distance from human judgments in open-source models.

Abstract

The rapid advancement of Large Language Models (LLMs) and their potential integration into autonomous driving systems necessitates understanding their moral decision-making capabilities. While our previous study examined four prominent LLMs using the Moral Machine experimental framework, the dynamic landscape of LLM development demands a more comprehensive analysis. Here, we evaluate moral judgments across 52 different LLMs, including multiple versions of proprietary models (GPT, Claude, Gemini) and open-source alternatives (Llama, Gemma), to assess their alignment with human moral preferences in autonomous driving scenarios. Using a conjoint analysis framework, we evaluated how closely LLM responses aligned with human preferences in ethical dilemmas and examined the effects of model size, updates, and architecture. Results showed that proprietary models and open-source models exceeding 10 billion parameters demonstrated relatively close alignment with human judgments, with a significant negative correlation between model size and distance from human judgments in open-source models. However, model updates did not consistently improve alignment with human preferences, and many LLMs showed excessive emphasis on specific ethical principles. These findings suggest that while increasing model size may naturally lead to more human-like moral judgments, practical implementation in autonomous driving systems requires careful consideration of the trade-off between judgment quality and computational efficiency. Our comprehensive analysis provides crucial insights for the ethical design of autonomous systems and highlights the importance of considering cultural contexts in AI moral decision-making.

Figures (6)

• Figure 1: Radar plots of moral preferences across different LLM families. AMCE values indicate preferences: Species ($+$: humans, $-$: pets), Social Value ($+$: high status, $-$: low status), Relation to AV ($+$: pedestrians, -: passengers), Number ($+$: more, $-$: fewer), Law ($+$: lawful, $-$: unlawful), Intervention ($+$: inaction, $-$: action), Gender ($+$: female, $-$: male), Fitness ($+$: fit, $-$: unfit/obese), Age ($+$: young, $-$: elderly). Gray-filled areas represent human preferences. Each subplot represents a different model family: (A) GPT-3.5, (B) GPT-4, (C) GPT-4o/o1, (D) Claude, (E) Gemini, (F) Llama, (G) Gemma, and (H) Other models.
• Figure 2: Distances between LLMs and human moral judgments across model families. Violin plots show the distribution of distances from human judgments for each model family, with individual models represented as points. Different colors indicate different model families. Model names are labeled.
• Figure 3: Comparison of moral judgment distances between proprietary and open-source models Violin plots with embedded box plots compare the distribution of distances from human judgments across three model categories: proprietary models, all open-source models, and large open-source models with parameters exceeding 10B. Individual models are represented as points.
• Figure 4: Relationship between model size and distance from human judgments in open-source models. Different model families are represented by different colors. Model names are labeled. Horizontal axis (model size in billion parameters) are represented in logarithmic scale.
• Figure 5: Temporal changes in distance from human judgments across proprietary model families. Different model families are represented by different colors. Model names are labeled.