Contribution to the handbook
Priority · Assessed 2026-10-28
The question specifies a consequential deliberation mechanism: candidate statements are generated, a personalized reward model predicts participants’ rankings, and a voting rule selects a statement. The training signals and intended prediction target matter to system design, while the packet also makes clear that prediction accuracy and fair representation are not established. This offers a concrete mechanism and a need for testing, without endorsing it for governance.
Does this challenge the goal?
None identified
The question asks about a particular preference-estimation mechanism; it does not contest the handbook’s goal or its open choices about governance authority.
The described mechanism is a study design, not evidence that preference prediction or ranked selection is appropriate for public decisions.
Why this category?
- Material system mechanism for aggregating participant preferences
- Consequential accuracy and representation limits call for evaluation
The supplied descriptions do not establish how accurately the model predicts each person’s preferences, how it handles conflicting critiques, or whether the selected statement represents participants fairly. The records share a source root.
This is an AI assessment of contribution to the goal, separate from whether the question has an answer.
- How did the AI mediator estimate which group statements people would prefer?
- The AI revised statements using feedback
- Selecting a group statement
- Two models with separate tasks
- Participants ranked and critiqued statements
- Participants rated and ranked statements
- Learning preferences
- Predicted rankings selected a winner
Making and interpreting laws →
Authority, accountability and resilience →
Tests, stop conditions and unresolved work →
Read the exploration goalDescribe the reward model, the participant rankings and ratings used to train it, and what its scores were intended to predict.
What we can say so far
The mediator used a personalized reward model: a 1.4-billion-parameter Chinchilla model with a linear layer that scored candidate statements. It was trained using participants’ statement rankings and quality ratings. Participants ranked statements by how much they agreed with them and separately rated argument quality, even when they disagreed with a statement. The scores were intended to predict how much each participant would agree with a candidate statement, not to establish that the prediction was accurate.
What remains open
- The supplied evidence does not establish the accuracy or reliability of the reward model’s predictions.
Conditions and limits
- This is the method reported in the supplied readable article text; supplementary materials are not included.
- The supplied evidence describes the model’s intended prediction, not how accurately it predicted preferences.
Objections to consider
- The method description does not establish how accurately or reliably the model predicted participants’ preferences. That limitation qualifies claims about performance but does not prevent describing the intended mechanism.
Evidence and related ideas
Supporting evidence, qualifications and context play different roles in this answer.
context
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.
Provides the broader training context and states that group approval was a goal, but does not detail the reward model’s training inputs.
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 →