Contribution to the handbook
Supporting · Assessed 2026-10-28
The evidence explains how the helper drafted and revised statements, its stated inspiration, and how successive rounds of human feedback were used to train it toward capturing shared views and group approval. This helps explain a possible deliberation mechanism, but does not establish that it advances governance or that approval is an adequate measure of citizens’ interests.
Does this challenge the goal?
In-scope gap
It asks about the design and training of a deliberation aid, which is relevant background for exploring ways citizens’ views could inform governance. It does not contest the exploration’s goal or propose that this particular method be adopted.
The evidence describes the study’s design and training, not whether they produce fair or reliable public input.
Why this category?
- system mechanism background
- limits of evidence
The cited records derive from the same article, not independent evidence. They describe the method and training objective; they do not establish that the helper captures every participant’s views or that group approval demonstrates fair representation.
This is an AI assessment of contribution to the goal, separate from whether the question has an answer.
Making and interpreting laws →
Authority, accountability and resilience →
Tests, stop conditions and unresolved work →
Read the exploration goalDescribe how it drafted statements from participants’ views and revised them in response to critiques, including its stated design inspiration.
What we can say so far
The article describes the Habermas Machine as drafting a group statement from participants’ opinions, then using their critiques to generate revisions. Across successive rounds, researchers used supervised fine-tuning and reward modeling to improve the system’s ability to capture shared views and maximize group-approval ratings. The stated intellectual inspiration was Jürgen Habermas’s theory of communicative action, including the idea that agreement may emerge through discussion under ideal conditions.
What remains open
- The supplied text does not specify how the system handled every kind of disagreement or critique.
Conditions and limits
- This is based on an AI interpretation of the readable article copy; supplementary materials are not included, and it is not independent verification.
- The article account does not explain how the system handled every kind of disagreement or critique.
Evidence and related ideas
Supporting evidence, qualifications and context play different roles in this answer.
supports
The AI revised statements using feedback The AI helper drafted a group statement from participants’ opinions, then revised it using their critiques. The design drew on Jürgen Habermas’s theory of communicative action; across successive rounds, researchers used supervised fine-tuning and reward modeling to improve the system, with group approval as the stated goal.
Directly describes drafting from participants’ opinions, revising from critiques, successive-round supervised fine-tuning and reward modeling, and Habermas-inspired design.
AI can help humans find common ground in democratic deliberation
Read full source: AI can help humans find common ground in democratic deliberation (opens in a new tab) ↗Read source discussion: AI can help humans find common ground in democratic deliberation →
context
A private group mediator The researchers developed an AI mediator that privately collects each participant’s views before proposing a statement intended to be acceptable to the group. They call this private, individual exchange caucus mediation.
Describe the successive-round human-data training, including supervised fine-tuning, reward modeling, and the goal of improving shared-view capture and group approval.
What we can say so far
According to the supplied article account, researchers gathered participant opinions and critiques over successive rounds. They fine-tuned a Chinchilla 70B model on material from earlier rounds, using highly rated group statements as targets. The supervised fine-tuned model generated candidate statements; a separate personalized reward model, also fine-tuned from Chinchilla, learned from participants’ rankings and quality ratings to estimate how much each person would agree with candidate statements. The stated training goal was to better capture shared views and maximize group approval ratings.
What remains open
- The supplied passages do not give a detailed round-by-round account of the training procedure or explain how every kind of disagreement or critique was handled.
Conditions and limits
- This is a description attributed to the supplied article account; supplementary materials are not included, and the account is not independent verification.
- The passages describe training design and goals, not whether the reward model accurately predicted preferences or whether training reliably improved group approval.
Objections to consider
- The passage cautions against conflating the AI’s training goal with observed changes in participants’ views. The answer describes the training goal only and does not claim that it caused those changes.
Evidence and related ideas
Supporting evidence, qualifications and context play different roles in this answer.
supports
The AI revised statements using feedback The AI helper drafted a group statement from participants’ opinions, then revised it using their critiques. The design drew on Jürgen Habermas’s theory of communicative action; across successive rounds, researchers used supervised fine-tuning and reward modeling to improve the system, with group approval as the stated goal.
Describes successive-round human feedback, supervised fine-tuning and reward modeling, and the stated shared-view and group-approval goals.
AI can help humans find common ground in democratic deliberation
Read full source: AI can help humans find common ground in democratic deliberation (opens in a new tab) ↗