Most AI is built for one person at a time: one user, one question, one answer. But many of the decisions that matter most are made in groups, where people disagree, talk past each other and slowly find a position everyone can live with.
In October 2024, researchers at Google DeepMind published a study in Science that tested whether AI could help with that. They called their system the Habermas Machine, after the philosopher who wrote about how reasonable people reach agreement through discussion.
How it worked
The system pairs two language models. The first writes group statements that try to reflect everyone's views. The second predicts how likely each participant is to accept a given statement. Group members wrote their opinions, the AI drafted a shared statement, participants critiqued it, and the AI revised.
The researchers tested it with 5,734 participants, including a demographically representative sample of the UK recruited through the Sortition Foundation.
What happened
- Participants preferred the AI's group statements over those written by human mediators 56% of the time.
- Agreement within groups rose by about eight percentage points on average.
- In a virtual citizens' assembly of 200 people, discussing topics from Brexit to universal childcare, the results held.
The researchers were careful about the limits. Groups did not simply drift toward the majority view, and the system does not fact-check or turn agreements into policy. As MIT Technology Review put it, it will not replace human mediators any time soon.
Why it matters for Social AI
The Habermas Machine worked with written opinions. Real deliberation is spoken, overlapping and emotional. People interrupt, hesitate, soften their tone or dig in. A mediator, human or machine, reads all of that.
To move from text-based experiments to AI that can support a live town hall, a workshop or a board meeting, models need to learn from recordings of real groups deliberating: who speaks, who gets cut off, when a room shifts. That data barely exists today, and it can only be collected with the consent of everyone in the room.
Crosstalk sources exactly that: consented, multi-party recordings of real people in conversation, from team workshops to panels and debates. Explore our datasets or read our paper, The Future Economics of Social AI.
Sources: MIT Technology Review, Singularity Hub.