How Debate Structure Reshapes AI Models' Moral Judgments

When large language models increasingly advise people on personal, mental health, and ethical matters, their ability to reason soundly about morality becomes critical. Yet most research on AI alignment has focused on isolated, single-turn responses. A new study from arXiv examines what happens when multiple models deliberate together over complex moral questions.

Studying Multi-Agent Moral Reasoning#

Researchers designed an experiment where they set three leading language models—GPT-4.1, Claude 3.7 Sonnet, and Gemini 2.0 Flash—to work through 1,000 real-world ethical dilemmas drawn from Reddit's "Am I the Asshole" community. Rather than relying on one-shot answers, they tested how these systems behave when collaborating to reach consensus on who bears responsibility in each scenario.

The key innovation was varying how the models interact. Some discussions happened synchronously, where all three models responded in parallel without hearing the others first. Others followed a round-robin format, where models took turns, hearing previous arguments before staking their own positions. This variation mirrors how multi-agent AI systems are actually deployed in practice.

Striking Differences in Flexibility and Values#

The results exposed major behavioral gaps across the models. GPT-4.1 proved remarkably stubborn, revising its initial judgments only 0.6–3.1% of the time, regardless of what others proposed. By contrast, Claude 3.7 Sonnet and Gemini 2.0 Flash showed far greater willingness to adjust, changing their verdicts between 28–41% of the time.

The models also revealed distinct value priorities. GPT-4.1 leaned toward personal autonomy and direct communication, weighting these principles more heavily than its peers. Claude 3.7 Sonnet and Gemini 2.0 Flash, meanwhile, emphasized empathetic and dialogue-based approaches to ethical reasoning.

The Power of Debate Format#

Perhaps most striking was how deeply the interaction structure shaped behavior. When models debated synchronously versus sequentially, their reasoning patterns shifted significantly. GPT-4.1 and Gemini 2.0 Flash proved highly sensitive to order effects—the sequencing of who speaks first had an outsized influence on their final positions. Claude 3.7 Sonnet remained more consistent across formats.