"Large-scale moral machine experiment on large language models"

Paper proposes a large-scale comparison of moral reasoning capabilities across major LLM families

Bigger models make more human-like moral choices in critical situations

https://arxiv.org/abs/2411.06790

🎯 Original Problem:

Evaluating moral decision-making capabilities of LLMs for autonomous driving systems, specifically understanding how different models handle ethical dilemmas in unavoidable accident scenarios.

-----

🛠️ Solution in this Paper:

→ Conducted extensive testing of 51 different LLMs including proprietary (GPT, Claude, Gemini) and open-source models (Llama, Gemma)

→ Generated comprehensive Moral Machine scenarios with six primary dimensions: species, social value, gender, age, fitness, and utilitarianism

→ Used conjoint analysis framework to analyze responses and compare with human preferences

→ Evaluated models through 50,000 scenarios (10,000 for some due to API constraints)

-----

💡 Key Insights:

→ Model size matters more than proprietary/open-source distinction for moral judgment quality

→ Open-source models >10B parameters perform similarly to proprietary models

→ Model updates don't consistently improve alignment with human preferences

→ Many LLMs show excessive emphasis on specific ethical principles

→ Significant negative correlation between model size and distance from human judgments

-----

📊 Results:

→ Proprietary models showed median distance of 0.9 from human judgments

→ Open-source models showed larger distances (median 1.2)

→ GPT-4 family demonstrated closest alignment with minimum distance of 0.6

→ Large-scale open-source models (>10B parameters) achieved 0.9 median distance