Contribution to the handbook
Priority · Assessed 2026-10-28
The question concerns both a consequential risk—different prediction accuracy across participant groups—and the test used to examine it. The evidence describes unequal-variance two-sample tests and training-data scope, and reports differences for several group comparisons while noting limits such as missing effect sizes or causes. These findings can inform evaluation of any proposed system that uses predictions to represent citizen preferences, without treating significance as proof of practical harm.
Does this challenge the goal?
In-scope gap
Testing whether preference-prediction accuracy varies across groups bears on citizen representation and evaluation within the existing goal; it does not challenge that goal.
The supplied findings concern the particular participants and models studied; they do not prove a general pattern of unequal representation or identify causes.
Why this category?
- Consequential group-representation risk
- Testing method and limits relevant to evaluation
These records derive from one study source. Statistical significance does not give the size or practical consequence of the gaps; the packet does not establish whether they recur in other populations or systems.
This is an AI assessment of contribution to the goal, separate from whether the question has an answer.
- How did the researchers test for accuracy differences between groups?
- Comparing participant groups
- Sex-group accuracy results
- Testing group accuracy
- Accuracy and bias varied
- Accuracy varied across groups
- Age and sex results varied
- Education and ideology differences
- Differences by political view and education
- How group accuracy was checked
- Age-group accuracy results
Authority, accountability and resilience →
Tests, stop conditions and unresolved work →
Read the exploration goalDescribe the group-comparison test, its unequal-variance allowance, and the training-data scope of the models tested.
What we can say so far
The researchers compared average accuracy between participant groups using two-sample t-tests that allowed unequal variances. The tests used models trained on half of the participant data. They also report retraining on balanced samples when the training data overrepresented a group, then testing accuracy separately for each group.
What remains open
- The supplied descriptions do not give group sizes, group-average accuracies, or effect sizes; those details are not needed to identify the test procedure.
Conditions and limits
- This describes the researchers’ reported methods, not independent verification of the underlying study.
- Balancing overrepresented groups does not by itself establish that all possible sources of bias were removed.
Evidence and related ideas
Supporting evidence, qualifications and context play different roles in this answer.
context
Comparing participant groups The authors plan to examine whether models predict choices differently across participants’ self-reported political ideology, age, education, and sex. For these comparisons, they balance the groups by downsampling the group with more participants; the supplied text does not include the results.
Describes planned comparisons across self-reported participant groups and downsampling, but does not specify the statistical test or training-data scope.
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) ↗Read source discussion: Large Language Models (LLMs) as Agents for Augmented Democracy →
supports
How group accuracy was checked To compare prediction accuracy across groups, the authors retrained models on balanced samples when the training data overrepresented one group, then tested accuracy separately for each group.
Adds that models were retrained on balanced samples when one group was overrepresented and tested separately by group.