Contribution to the handbook
Supporting · Assessed 2026-10-28
The study setup and validation limits provide useful background for judging claims about AI preference prediction. The evidence says the data came from an online participatory experiment and cautions that matching its participants does not establish representativeness or performance elsewhere; it does not specify a governance design or demonstrate deployment capability.
Does this challenge the goal?
None identified
The question seeks details about a study’s data and models; it does not challenge the exploration’s goal or scope.
The study details do not establish how preference-prediction methods would perform in a real governance system.
Why this category?
- useful background on data provenance and representativeness limits
- relevant limits on interpreting model validation
The supplied evidence derives from one study and shared source material; it does not establish broader population validity, real-world performance, or the adequacy of the reported categories.
This is an AI assessment of contribution to the goal, separate from whether the question has an answer.
- What data and models did the authors study?
- Models in the comparison
- Demographic details and ethics review
- The data used in the comparison
- Testing individual predictions
- Adapting the models
- Six fine-tuned models
- No new human data collected
- Predictions improved some sample matches
- Participant details used as context
- Keeping model architecture consistent
- Where the data came from
- Six models tested on collected preferences
- Participants compared policy proposals
Authority, accountability and resilience →
Tests, stop conditions and unresolved work →
Read the exploration goalNew evidence is being assessed. Any answer below is provisional; parts whose evidence changed or was withdrawn are withheld.
Identify both the study data and the models tested by the authors.
What we can say so far
The authors analyzed participant choices between pairs of policy proposals, along with demographic information. The data came from Brazucracia, an online government-program builder used during Brazil’s 2022 presidential election; the paper reports that it collected no new primary data from human participants. The authors fine-tuned six language models on these preference data: GPT-3.5 Turbo, LLaMA-2 7B, LLaMA-3 8B, Mistral 7B, Falcon 7B, and Gemma 7B. They also compared them with BERT SQuAD as a simpler benchmark. The models were tested on predicting participants’ choices and group preferences. These are the data and models described for this study, not evidence that its participants represented all Brazilians or that the findings generalize beyond this setting.
What remains open
No material gap is recorded within this part’s current scope.
Conditions and limits
- The participant data originated in an online project during Brazil’s 2022 presidential election; the paper reports no original human-participant research or primary data collection by this study.
- The online participant choices should not be treated as representative of all Brazilians.
- The model list is limited to those included in the reported comparison; GPT-4 and Claude 3 were not included because fine-tuning options were not publicly available to the authors at the time.
Objections to consider
- The comparison was not exhaustive of available language models; the answer identifies the models included and does not imply that excluded models were tested.
- The data came from online participants and are not established as representative of the broader Brazilian public. The answer reports the source and setting without making a representativeness claim.
Evidence and related ideas
Supporting evidence, qualifications and context play different roles in this answer.
supports
No new human data collected The authors say the study involved no original research with human participants and collected no primary data.
Supports the distinction that this study collected no new primary human-participant data.
Large Language Models (LLMs) as Agents for Augmented Democracy
Read full source: Large Language Models (LLMs) as Agents for Augmented Democracy (opens in a new tab) ↗