Contribution to the handbook
Supporting · Assessed 2026-10-28
Question generation, review, and topic selection offer relevant background on structuring deliberation and reducing discussion risks. The evidence describes model-generated candidates and human review, but does not establish that screening removed harmful content or that the process would work in other settings (s-habermas-machine-d0f63279c1aef658; s-habermas-machine-e715435e78010161; s-habermas-machine-ab39b98ffd3e4f12).
Does this challenge the goal?
None identified
The question concerns how a study prepared discussion questions, not whether the handbook’s goal or scope should change.
The evidence describes the study’s process but does not establish how well it would transfer to civic governance.
Why this category?
- Provides a potentially useful deliberation-design example and a limited safeguard; the effectiveness and transferability of the process are not established.
The packet does not report review outcomes, how reviewers resolved disagreements, or whether the process reduced risk in practice.
This is an AI assessment of contribution to the goal, separate from whether the question has an answer.
- How did the authors organize discussion questions into topic groups?
- How did the authors create and review the discussion questions?
- How did researchers prepare discussion questions?
- Groups shared topics chosen as important
- Safeguards can guide discussion
- Generating and screening questions
- Questions were generated and reviewed
Making and interpreting laws →
Authority, accountability and resilience →
Read the exploration goalNew evidence is being assessed. Any answer below is provisional; parts whose evidence changed or was withdrawn are withheld.
Explain how candidate questions were generated, reviewed, and assembled, without equating screening with elimination of offensive content.
What we can say so far
The authors used Chinchilla 70B to generate 10,000 candidate discussion questions from 175 author-written examples. They reviewed them for safety, ethics, clarity, and whether they could support a range of views, with screening intended to reduce the risk of offensive comments—not shown to eliminate that risk. They assembled about 5,500 distinct questions. Separately, all virtual-assembly groups discussed the same nine questions, chosen because participants had rated the topics as highly important during training.
What remains open
- The material does not report review outcomes, how many questions were removed, the detailed review criteria, or how reviewers resolved disagreements.
Conditions and limits
- The account is based on interpretations of the readable article text; supplementary materials are not included, and the interpretations are not independent evidence.
- The review is described as a process and risk-reduction aim, not as proof that every question was free of risk or bias.
Evidence and related ideas
Supporting evidence, qualifications and context play different roles in this answer.
supports
Generating and screening questions Researchers prompted a language model with a small set of examples to generate candidate discussion questions. They reviewed the questions to reduce the risk of offensive comments and assembled about 5,500 distinct questions.
Supports the screening aim and assembly of about 5,500 distinct questions; cautions that selection of final questions for each group is not described in that passage.
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 →
supports
Groups shared topics chosen as important All groups in the virtual assembly discussed the same nine questions, chosen because participants had rated the topics as highly important during training. The assembly took place over three weekly sessions, each lasting one hour.