Seminar: Utilitarian Boosts in Group Moral Judgment: Humans and Multi-Agent LLMs

In this talk, I first present experimental work on human moral judgment in small groups and then turn to analogous phenomena in multi-agent large language models (LLMs). On the human side, we study how social interaction shapes moral evaluations of real-life and sacrificial dilemmas in which an agent’s action or inaction violates a moral rule to benefit the greater number. Groups of 4–5 participants judge these dilemmas first individually and privately, then collectively through face-to-face or online discussion, and finally individually again. Collective judgments are more utilitarian than the average of their members’ individual judgments, revealing a “utilitarian boost” under group deliberation. In a second experiment, we also track state anxiety before, during, and after online interaction, and find that the boost in collective utilitarian judgments is accompanied by reduced anxiety, consistent with the idea that social interaction transiently lowers the emotional cost of endorsing norm violations. I then turn to LLM-generated moral judgments and ask whether a comparable pattern emerges when models reason in small “groups.” Using six models evaluated on established sets of moral dilemmas, I compare a Solo condition, where models respond independently, to Group conditions, where pairs or triads engage in multi-turn exchanges before producing a joint recommendation. In personal sacrificial dilemmas involving direct harm for the greater good, all models exhibit more utilitarian recommendations in the Group setting. However, response decompositions indicate a different functional profile than in humans: group-induced changes in LLM outputs are better captured by reduced weighting of norm-violation cues or more impartial treatment of beneficiaries. I close by discussing model differences, how prompt design and agent composition can amplify or dampen these multi-agent effects, and the implications for AI alignment, multi-agent system design, and artificial moral reasoning.
Dr Anita Keshmirian is a multidisciplinary cognitive scientist from Iran, based in Berlin. She is Head of Data Science and Assistant Professor of Data Science and Psychology at Forward College, a visiting scientist at the Human-Centered AI division of Helmholtz Munich, and a guest researcher at the Faculty of Electrical Engineering and Computer Science at TU Berlin. Trained in computer engineering, cognitive science, and neuroscience, she obtained her PhD from LMU Munich on moral decisions in and for groups. Previously, she worked at Harvard, Charité Berlin, LMU Munich, and Fraunhofer IKS. Her research investigates how individual and collective moral and causal judgments arise in humans and AI systems, using behavioral experiments, Bayesian networks, and multi-agent LLMs.

Alternatively, join online:

Microsoft Teams Meeting ID: 378 283 009 672 8 Passcode: 5yK65cy7