Large Language Models Amplify Human Biases in Moral Decision-Making

As large language models (LLMs) become more widely used, people increasingly rely on them to make or advise on moral decisions. Some researchers even propose using LLMs as participants in psychology experiments. It is therefore important to understand how well LLMs make moral decisions and how they compare to humans. We investigated this question in realistic moral dilemmas using prompts where GPT-4, Llama 3, and Claude 3 give advice and where they emulate a research participant. In Study 1, we compared responses from LLMs to a representative US sample (N = 285) for 22 dilemmas: social dilemmas that pitted self-interest against the greater good, and moral dilemmas that pitted utilitarian cost-benefit reasoning against deontological rules. In social dilemmas, LLMs were more altruistic than participants. In moral dilemmas, LLMs exhibited stronger omission bias than participants: they usually endorsed inaction over action. In Study 2 (N = 490, preregistered), we replicated this omission bias and document an additional bias: unlike humans, LLMs (except GPT-4o) tended to answer ``no'' in moral dilemmas, whereby the phrasing of the question influences the decision even when physical action remains the same. Our findings show that LLM moral decision-making amplifies human biases and introduces potentially problematic biases.