Skip to content
Citizen Infrastructure

AI mediation & the Habermas Machine

Group deliberation works, but it’s expensive and hard to scale — and good human mediators are scarce. So a DeepMind team (Michael Henry Tessler, Michiel Bakker and colleagues) asked a sharp question: can an AI play the mediator, helping a divided group find a statement they can all live with? Their system, published in Science in October 2024, is the Habermas Machine — named after Jürgen Habermas, a founding thinker of deliberative democracy.

How it works: caucus mediation

Section titled “How it works: caucus mediation”

The design is deliberately modest. A small group (in the study, 3–5 people) each answers a question — say, should the government provide universal free childcare? — by writing their opinion privately; they don’t see each other’s. The AI then:

  1. Generates many candidate “group statements” from those opinions.
  2. Ranks each candidate for each participant, using a model that predicts how much a person holding a given opinion would endorse a given statement.
  3. Aggregates those personal rankings with a social-choice (voting) rule, and picks the statement that wins — like running a tiny virtual election.

Participants can then critique the draft, and the machine revises it. Because the whole thing is a ranking, it’s auditable: you could in principle see where any statement sat in your own preference order. This is “caucus mediation” — people interact with the mediator, not directly with each other.

What they found

Section titled “What they found”
  • Preferred over humans. Against paid, trained, financially-incentivised human mediators, the AI’s statements were chosen more often — judged clearer, more informative, and fairer in how they aggregated views.
  • Less division. Groups measurably converged: agreement rose after deliberation, and the effect came from the AI’s consensus statement, not merely from seeing others’ opinions.
  • Minorities kept. The statement weighted minority opinions roughly in proportion to their share (≈30% of opinions → ≈30% of the statement), and critique increased that without alienating the majority — a guard against majority tyranny.
  • A real assembly. With the Sortition Foundation, 200 representative UK citizens ran a virtual citizens’ assembly over nine divisive questions and showed the convergence Habermas himself predicted.

The honest limits

Section titled “The honest limits”

The authors call it “minimally deliberative.” People don’t talk directly, there’s no fact-checking or learning phase, rounds are short (15–20 minutes), and it was only tested in groups up to five. On calcified questions (Brexit) it moved nothing, and average belief shifts were small. Critics at the launch pressed two worries worth keeping: aggregating opinions isn’t the same as people reasoning together (real deliberation), and a group finding “common ground” can also drift toward a more extreme shared position. The fair reading: this is one component — assistive civic AI, in the eyeglasses-not-replacement sense — that could slot into a larger deliberative process, not a replacement for it.

Sources

Section titled “Sources”