AI Mediation Tool Helps Find Common Ground in Group Debates

Reaching a consensus in group deliberations is challenging, especially when participants hold widely varied social, political, and ideological views. But a recent study by Google DeepMind suggests that AI could aid in finding common ground.

Researchers at Google DeepMind have developed a tool using large language models (LLMs) as a “caucus mediator” to summarize areas of agreement within group discussions. This tool, called the Habermas Machine (HM)—inspired by the German philosopher Jürgen Habermas—aims to enhance understanding without replacing human mediators.

“The large language model was trained to identify and present areas of overlap between the ideas held among group members,” says Michael Henry Tessler, a research scientist at Google DeepMind. “It was not trained to be persuasive but to act as a mediator.” The study is published in Science.

How It Works: Google DeepMind’s Experiment

Google DeepMind recruited 5,734 participants, some through crowdsourcing and others through the Sortition Foundation, a nonprofit organizing citizen assemblies, to test HM. The Sortition Foundation participants represented a demographically balanced sample of the UK population.

The Model’s Structure: The HM uses two LLMs fine-tuned for mediation. The first model generates statements to reflect diverse views within the group. The second, a reward model, scores the statements based on predicted group agreement levels.

Testing the HM in Groups: Participants answered questions like “Should we lower the voting age to 16?” and “Should the National Health Service be privatized?” Groups of five discussed their responses after submitting them to the HM, which then generated summaries for individuals to review and critique. Final statements were ranked by participants.

Human vs. AI Mediation: How the Habermas Machine Performed

To evaluate the AI’s effectiveness as a mediator, the study divided participants into six-person groups. In each group, one person served as a human mediator, writing statements on behalf of the group. Simultaneously, the HM generated an AI-mediated statement. Participants then selected their preferred statement.

Results: Over half of the participants (56%) chose the AI-generated statements, citing higher quality and stronger alignment with group opinions. This led to less division in participant views post-deliberation.

Limitations and Ethical Concerns

While promising, the AI mediation tool has notable limitations. According to Joongi Shin, a researcher at Aalto University studying generative AI, transparency is crucial: “Unless the situation or the context is very clearly open, so they can see the information that was inputted into the system and not just the summaries it produces, I think these kinds of systems could cause ethical issues.”

Additionally, Google DeepMind did not explicitly inform participants in the experiment that an AI system would be generating opinion summaries. Participants consented with an acknowledgment of algorithm involvement, but AI’s specific role was not fully disclosed.

Michael Henry Tessler of DeepMind adds that the model, in its current form, lacks certain essential mediation abilities. “For example, it doesn’t have the mediation-relevant capacities of fact-checking, staying on topic, or moderating the discourse,” he says. Google DeepMind has no immediate plans to release the model publicly and acknowledges that further study is necessary to understand its best applications.

Future Prospects for AI in Mediation

The Habermas Machine demonstrates AI's potential in supporting group consensus by presenting neutral, well-rounded summaries that aid in reducing divisiveness. While promising, AI mediation faces challenges: refining capabilities for fact-checking, keeping discussions on topic, and addressing ethical concerns around transparency. These are areas where human oversight remains essential. Further research will be key to ensuring responsible AI use and understanding where this technology could best support civic and political discussions.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.