The moral machine experiment on large language models.

Area of Science:

• Artificial Intelligence Ethics
• Human-Computer Interaction
• Autonomous Systems Morality

Background:

• Large language models (LLMs) are increasingly integrated into critical sectors, necessitating an understanding of their ethical decision-making.
• Autonomous driving systems require robust ethical frameworks to navigate complex moral dilemmas.

Purpose of the Study:

• To investigate the moral judgment tendencies of prominent LLMs using the Moral Machine framework.
• To compare LLM ethical decision-making with established human preferences in simulated accident scenarios.
• To identify potential discrepancies and similarities between LLM and human moral reasoning for autonomous driving applications.

Main Methods:

• Utilized the Moral Machine framework to present ethical dilemmas to various LLMs.
• Collected and analyzed decision-making data from GPT-3.5, GPT-4, PaLM 2, and Llama 2.
• Compared LLM responses against a large dataset of human preferences.

Main Results:

• LLMs and humans generally align on prioritizing human lives over animals and saving more individuals.
• PaLM 2 and Llama 2 demonstrated notable deviations from human moral preferences.
• Significant quantitative differences were observed, with LLMs potentially making more absolute judgments than humans.

Conclusions:

• LLMs exhibit both alignment and divergence from human moral judgments in autonomous driving contexts.
• Specific LLMs like PaLM 2 and Llama 2 require further ethical refinement for safety-critical applications.
• Understanding these ethical frameworks is crucial for the responsible development and deployment of AI in autonomous vehicles.