Contribution to the handbook
Supporting · Assessed 2026-10-28
The two-model arrangement is useful background for understanding one AI-assisted statement-generation process, but the supplied evidence does not demonstrate governance performance or establish that these components should be adopted.
Does this challenge the goal?
None identified
The question asks about the architecture of a study’s deliberation system; it does not contest the exploration’s goal, boundary, or assumptions.
The supplied evidence describes the study setup but does not establish that this architecture is suitable for public governance.
Why this category?
- The evidence distinguishes a model that generates candidate statements from a personalized reward model that predicts preferences, while explicitly not establishing prediction accuracy (s-habermas-machine-302a65acf0fbb34d).
- Training data and signals are described, but their quality, representativeness, and resulting model performance are not established in the supplied evidence (s-habermas-machine-3aff75bad42df435; s-habermas-machine-5be6d6107b0adf5a).
These are interpretations of the same article, not independent evidence. The packet does not establish accuracy, safety, or transferability to civic decision-making.
This is an AI assessment of contribution to the goal, separate from whether the question has an answer.
Making and interpreting laws →
Tests, stop conditions and unresolved work →
Read the exploration goalIdentify the two fine-tuned Chinchilla models and their separate roles: generating candidate statements and predicting individuals’ preferences among them.
What we can say so far
The system used two fine-tuned Chinchilla models: a supervised fine-tuned model generated candidate statements, and a personalized reward model predicted each person’s preferences among those statements. The reward model’s design and training signals are described in more detail, but the supplied material does not establish its prediction accuracy.
What remains open
No material gap is recorded within this part’s current scope.
Conditions and limits
- This is a description of the reported setup, not evidence that the generated statements were good or preference predictions accurate.
- The source is an AI interpretation of a readable copy of the article; supplementary materials are not included, and the evidence is not independent verification.
Evidence and related ideas
Supporting evidence, qualifications and context play different roles in this answer.
supports
Two models with separate tasks The system has two models. A supervised fine-tuned model generates candidate statements, and a personalized reward model predicts each person’s preferences among them. Both were fine-tuned versions of Chinchilla.
Directly identifies both fine-tuned Chinchilla models and their separate roles.
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
Training material The authors fine-tuned a Chinchilla 70B language model on opinions and critiques from earlier rounds, using highly rated group statements as targets. The material came from 578 groups, 2,473 participants, and 1,630 questions.
Adds reported training material and sample counts, but does not independently establish the two-model role distinction.
AI can help humans find common ground in democratic deliberation
Read full source ↗