Large language models show amplified cognitive biases in moral decision-making

Area of Science:

• Cognitive Science
• Artificial Intelligence Ethics
• Moral Psychology

Background:

• Increasing reliance on large language models (LLMs) for decision-making and advice.
• Growing interest in using LLMs as participants in psychological research.
• Need to understand LLM moral decision-making capabilities and compare them to human behavior.

Purpose of the Study:

• To investigate how LLMs make moral decisions and advise on moral dilemmas.
• To compare LLM moral decision-making with human responses across various dilemma types.
• To identify and analyze biases present in LLM moral judgments.

Main Methods:

• LLMs were prompted to emulate or advise on decisions in realistic moral dilemmas.
• LLM responses were compared to human participant data (N=285, 474, 491) across four studies.
• Dilemmas included collective action problems and moral scenarios contrasting utilitarian and deontological reasoning.
• LLMs with and without fine-tuning were compared to identify bias origins.

Main Results:

• LLMs demonstrated greater altruism than humans in collective action problems.
• LLMs exhibited a stronger omission bias, favoring inaction over action in moral dilemmas.
• LLMs showed a response-conditional bias, altering decisions based on question wording.
• These biases were replicated across different dilemma sets and found to likely stem from chatbot fine-tuning.

Conclusions:

• LLMs possess distinct moral decision-making patterns compared to humans, including amplified omission and response-conditional biases.
• Fine-tuning for chatbot applications appears to be a significant source of these LLM biases.
• Uncritical adoption of LLM moral advice risks exacerbating existing human biases and introducing new problematic ones.