Contribution to the handbook
Supporting · Assessed 2026-10-28
Training-data provenance is useful background for scrutinizing preference models, but the question is narrowly about one study’s training materials. The supplied evidence describes fine-tuning inputs, target statements, data scale, and public availability, while noting that these details do not establish model quality or representativeness (s-habermas-machine-3aff75bad42df435; s-habermas-machine-f3d49a0fb972ca97).
Does this challenge the goal?
None identified
The question asks for a study-method detail; it does not contest the handbook’s goal, beneficiary boundary, assumptions, or desired outcome.
The question does not itself state how the training method should affect a governance design.
Why this category?
- Useful background on model provenance and data transparency; no direct system specification or demonstrated governance implication.
The packet covers the article text but not supplementary materials; it does not establish the data’s representativeness or the models’ performance.
This is an AI assessment of contribution to the goal, separate from whether the question has an answer.
Authority, accountability and resilience →
Tests, stop conditions and unresolved work →
Read the exploration goalIdentify the model’s reported fine-tuning inputs and target statements, and the source and scale of that material.
What we can say so far
The article reports that a 70-billion-parameter Chinchilla language model was fine-tuned using opinions and critiques from earlier rounds, with highly rated group statements as target outputs. The reported material came from 578 groups, 2,473 participants, and 1,630 questions. The system also included a separate 1.4-billion-parameter Chinchilla reward model, trained on participants’ rankings of candidate statements and quality ratings; its scores were intended to estimate each participant’s expected agreement. The article says the question and human-feedback dataset used for training and evaluation is publicly available, with the seed questions in supplementary materials.
What remains open
- The supplied passages do not assess the quality or representativeness of the training examples.
- The supplementary materials are not included in the supplied evidence, so further details available only there cannot be assessed.
Conditions and limits
- These are descriptions of the authors’ reported training setup, not independent verification of the underlying materials.
- The supplied copy includes the article but not its supplementary materials.
- The reported counts and training choices do not establish the quality or representativeness of the examples, or the models’ performance.
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 using participants’ views to draft statements and their critiques to refine them across rounds.
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 →