POSTPONED: Anita Keshmirian (Forward College, Berlin), “Many minds, diverging morals: Human groups vs. AI in moral decision-making”

Abstract

“Moral judgments are inherently social, shaped by interactions with others in everyday life. Despite this, psychological research has rarely examined the impact of social interactions on these judgments. In our study, we explored the role of group dynamics in moral decision making by having small groups (4-5 participants) evaluate moral dilemmas first individually, then collectively, and finally individually a second time. Participants judged real-life and sacrificial moral dilemmas involving actions or inactions violating moral principles to benefit the greater good. Experiment 1 found that collective judgments were more utilitarian than individual judgments, supporting the hypothesis that group deliberation temporarily reduces the emotional burden of violating moral norms. Experiment 2 measured participants’ state anxiety and moral judgments before, during, and after online interactions. Results again showed that collectives were more utilitarian, reducing state anxiety during and after social interaction, suggesting that stress reduction may explain the shift toward utilitarianism in group settings.

We replicated this experiment using multi-agent large language models (LLMs) to test how artificial agents make moral decisions. Preliminary findings revealed that, unlike humans, groups of LLM agents were less utilitarian than individual agents. Analysis of the agents’ interactions showed a consistent pattern of virtue-signaling, with LLMs emphasizing deontological reasoning (focusing on moral rules) rather than utilitarian principles. This divergence from human behavior suggests that collective reasoning in AI systems is shaped by different dynamics, likely due to how LLMs are trained to prioritize socially accepted norms. These results highlight important differences in moral decision-making between human and artificial intelligence, offering new insights into the development of AI systems that more closely mirror human ethical reasoning, particularly in complex, real-world collective decision-making scenarios.”

Image credit: ©SCIoI/ generated with DALL-E