A new study shows that an artificial intelligence (AI) tool can help people with different views find common ground by more effectively summarizing the collective opinion of the group than humans. By producing statements that convey the majority opinion, while incorporating the minority's perspective, the AI produced outputs that participants preferred—and that they rated as more informative, clear, and unbiased, compared to those written by human mediators. Human society is enriched by a plurality of viewpoints, but agreement is a prerequisite for people to act collectively. Finding agreement through a free exchange of views is often difficult. Collective deliberation can be slow, difficult to scale, and not equally attentive to different voices. Inspired by recent demonstrations that showed that large language models (LLMs) can effectively summarize varying perspectives on a public debate platform, Michael Tessler and Michiel Bakker of Google DeepMind, and colleagues, explored whether an AI mediator could support the process of deliberation, helping people who disagree feel less divided, feel their viewpoint was heard, and find common ground.
In their United Kingdom-based study – involving several thousand people found on a crowdsourcing research platform – the authors asked human participants to write their opinions on complex and important topics. In an iterative process, participants, who were organized into small groups, submitted their personal opinions to an LLM that had been trained to generate a "group statement" – one focused on elevating shared viewpoints. Group statements were also written by human mediators. Participants could review both AI- and human-generated group statements and consider whether the statements aligned with their own views. The AI mediator's statements were preferred by participants 56% of the time vs. 44% for the human mediators' statements. After deliberating with the AI mediator, the small groups of participants were less divided in their positions on the issues, suggesting that meaningful common ground was found. The authors report the AI-generated statements were successful because they incorporated dissenting voices while respecting the majority position. The authors replicated their results in a real-world sample of UK participants that more closely reflected the diversity in the UK population, finding AI statements were highly endorsed and left groups less divided. "… the key translational opportunity provided [by this AI tool] …is that it facilitates collective deliberation that is time-efficient, fair and scalable," say the authors. Their new approach to collective deliberation circumvents some of the limitations of in-person debate, they say, while also noting it may miss out on other advantages that arise from in-person discussion, including the opportunity to build interpersonal relationships with other discussants.
