Finding agreement in democratic debate is hard when people begin from different ideological, political, and social positions. A Google DeepMind research project suggests that AI mediation may be able to help groups identify shared ground, at least in carefully designed deliberation settings.
The system, called the Habermas machine, was built to generate statements that reflect areas of overlap among participants. It was tested on contentious questions, including whether the voting age should be lowered to 16 and whether the National Health Service should be privatized.
What the Habermas machine was designed to do
Researchers from Google DeepMind trained a system of large language models to work as a “caucus mediator.” Its role was not to argue people into changing their minds. Instead, it was trained to identify common ground and produce summaries that could represent the group’s views.
The tool is named after the German philosopher Jürgen Habermas. Michael Henry Tessler, a research scientist at Google DeepMind, said the large language model was trained to find overlap between the ideas held by group members. He also stressed that it was built to act as a mediator, not as a persuasive system.
The Habermas machine has two parts. The first is a generative model that proposes statements reflecting the range of views in the group. The second is a personalized reward model that scores those proposed statements by estimating how much each participant would agree with them.
How Google DeepMind tested AI deliberation
Google DeepMind recruited 5,734 participants. Some came through a crowdsourcing research platform, while others came through the Sortition Foundation, a nonprofit that organizes citizens’ assemblies. The Sortition groups formed a demographically representative sample of the UK population.
The researchers tested the system in two steps. First, they examined whether it could accurately summarize collective opinions. Then they tested whether it could mediate between groups and help participants move toward common ground.
In the first stage, participants answered questions before discussing their views in groups of around five people. The Habermas machine then summarized the group’s opinions. Those summaries were sent back to individuals for critique, after which the system produced a final set of statements for participants to rank.
This design matters because the task was not simply to generate fluent text. The system had to take multiple viewpoints, capture agreement where it existed, and produce statements that people in the group could recognize as fair enough to support or critique.
Where the AI statements outperformed human mediators
The researchers then tested whether the Habermas machine could work as an AI mediation tool. Participants were placed into six-person groups. In each group, one participant was randomly assigned to write statements on behalf of everyone and was designated the “mediator.”
During each round of deliberation, participants saw two statements: one written by the human mediator and one generated by the Habermas machine. They then chose which statement they preferred.
The result was notable. More than half (56%) of the time, participants chose the AI statement. They judged the AI-generated statements to be higher quality than those written by the human mediator and tended to endorse them more strongly.
The groups also became less divided in their positions after deliberating with help from the AI mediator. That does not mean the technology resolved every disagreement. It does suggest that, under the conditions of the study, AI-generated summaries helped participants recognize shared positions more effectively than a randomly assigned human mediator.
The limits are as important as the results
The research shows promise for AI mediation, but the source also makes clear that the technology has significant limits. Joongi Shin, a researcher at Aalto University who studies generative AI, warned that systems like this could raise ethical issues if the situation is not open enough for people to see the information that went into the system, rather than only the summaries it produces.
Transparency is especially important because Google DeepMind did not explicitly tell participants in the human mediator experiment that an AI system would generate group opinion statements. The consent form did indicate that algorithms would be involved.
Tessler also described limits in the model’s current form. He said it does not have mediation-relevant capacities such as fact-checking, staying on topic, or moderating the discourse. Those are central functions in real-world deliberation, especially when discussions involve disputed facts, emotional conflict, or people talking past one another.
That distinction keeps the findings in perspective. The Habermas machine may be useful at summarizing viewpoints and surfacing agreement. It is not presented as a replacement for human mediators, and it is not ready to manage the full complexity of public deliberation on its own.
What this means for the future of AI mediation
The study points to a narrow but important possibility: large language models may help people see the parts of a disagreement where overlap already exists. In polarized discussions, that can be valuable because participants may not naturally notice shared assumptions or compatible goals.
At the same time, the future of tools like the Habermas machine depends on where and how they are used. The company says more research would be needed to support responsible and safe deployment. It also says it has no plans to launch the model publicly.
For now, the clearest lesson is that AI can assist with one part of deliberation: turning a set of different views into statements that people may find acceptable. The harder question is whether such systems can be made transparent, accountable, and useful outside a research setting.