Large Language Models (LLMs) as Agents for Augmented Democracy

What the idea means

A personalized agent, described as a digital twin, is software intended to reflect an individual’s preferences. The authors suggest it could assist that person’s participation.

Why it matters

This is a proposal for AI assistance, not for AI to rule or make political choices. The passages do not show that such an agent works in practice.

Question this helps explore

What role do the authors propose for voter-focused AI helpers?

Evidence and limits

The paper presents personalized agents as a possible use of language models. Its tests concern predictions from collected data, not whether an agent represents a voter in real decisions.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 2

What the idea means

The recorded choices concern specific proposals from the candidates’ programs, gathered in an experiment about building a government program together. The authors use this open dataset to study predictions.

Why it matters

The setting gives the authors policy-level choices to analyze, but the passages do not establish how well results apply beyond these volunteers, policies, and election.

Question this helps explore

What data and study setting do the authors use?

Evidence and limits

The experiment covered 67 policies. Volunteers chose between pairs and reported characteristics such as age, location, sex, and political orientation.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 2, 3

What the idea means

Fine-tuning means further training a model on particular data. Here, the authors adapt existing models using the volunteers’ recorded choices.

Why it matters

This describes the tested setup, not evidence that these models are suitable for political decisions in general.

Question this helps explore

Which models do the authors test, and how do they use them?

Evidence and limits

They adapt the models using collected policy-preference data, then use them to predict choices and group preferences.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 3

What the idea means

In this test, the models predicted choices better than assuming a person always prefers policies from the candidate matching their political orientation.

Why it matters

The result suggests detailed policy choices may contain information not captured by that broad political label. The passages give no accuracy figures and do not establish performance elsewhere.

Question this helps explore

How did the models compare with the political-alignment rule for individual choices?

Evidence and limits

The authors use train-test cross-validation, which evaluates predictions on data held aside from model fitting. The comparison is with this specific rule, not every possible alternative.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 1, 3

What the idea means

The authors compare an estimate based on a probability-based sample alone with one that also uses model predictions. They report that the augmented estimate was more accurate in their test.

Why it matters

This suggests a possible way to improve estimates from preference data. It does not establish that the method works for other populations or can replace careful sampling.

Question this helps explore

Did model predictions improve the group-preference estimate in the reported comparison?

Evidence and limits

The paper estimates accuracy for individual choices and group preferences. The passages state the direction of the group-level result, but not its size or uncertainty.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 1, 3

What the idea means

A person’s choices on specific policies may not always follow proposals associated with a broad political label. The authors say the models capture some of this detail in the study data.

Why it matters

This supports further research using detailed preferences, but does not prove that a model fully understands a person or can represent them fairly.

Question this helps explore

What do the authors infer from the individual prediction result?

Evidence and limits

This is the authors’ interpretation of models outperforming the rule based on self-reported political orientation. It does not show that models capture every aspect of a person’s views.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 1, 3

What the idea means

People could express views on separate proposals rather than only choosing a representative who offers a package of policies. The idea combines aspects of representative and direct participation.

Why it matters

This describes a way of expressing preferences; it does not necessarily propose replacing representative government, and the passages do not show a functioning system.

Question this helps explore

What does the paper mean by augmented democracy?

Evidence and limits

The approach lets people indicate preferences on individual policy proposals, with software agents proposed as a way to support detailed participation.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 4

What the idea means

Choosing a candidate means choosing among packages of proposals, which reduces the number of options. Parties may adjust their packages to voters, but a package is not guaranteed to match each person’s preferences.

Why it matters

This motivates exploring ways to express preferences on separate policies. The passage does not measure how often packages fail to fit voters.

Question this helps explore

Why might candidate policy packages fail to match every person’s views?

Evidence and limits

The authors say bundling reduces the number of choices citizens face. They also acknowledge incentives to adapt packages, while noting that this does not ensure a best fit for all.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 4

What the idea means

Agents could help people engage with individual policy choices rather than only choosing among candidate packages. The study does not show that agents reduce effort or improve participation in practice.

Why it matters

This is a possible use for AI assistance. Whether people would find it easier or trust an agent’s predictions remains open.

Question this helps explore

How might agents support detailed participation?

Evidence and limits

The paper presents this as an opportunity to explore, not as a tested solution to the effort involved.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 4

What the idea means

The authors say that many democracies place intermediaries between the public and political power. They connect that arrangement to recurring questions about whether people can participate more directly.

Why it matters

This helps explain why the paper explores other forms of participation, but does not establish that agents should replace existing intermediaries.

Question this helps explore

Why does the paper focus on alternatives to political intermediation?

Evidence and limits

This is the authors’ framing of the motivation for exploring participation tools, not a measured finding about how many democracies rely on intermediaries.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 2

What the idea means

The paper presents its proposal as part of a broader debate about whether technology can let people participate more directly, rather than relying only on political intermediaries.

Why it matters

This background explains why the authors examine AI assistance, but does not show that earlier approaches or AI solve the problem.

Question this helps explore

What related debates do the authors identify?

Evidence and limits

The authors present these developments as background for their study, not as findings from their own experiment.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 2

What the idea means

Democratic systems make choices about how much participation to enable and how people can take part. The authors connect this challenge to questions about political intermediaries.

Why it matters

This context explains why the paper explores participation tools, but does not show that technology resolves these trade-offs.

Question this helps explore

What broader problem motivates the search for new ways to participate?

Evidence and limits

These points are the paper’s framing, not results measured by the authors.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 2

What the idea means

The paper connects its proposal to a past vision of people and computers cooperating on decisions. It does not test whether this cooperation performs better than either working alone.

Why it matters

This context helps explain the interest in computer assistance, but is not evidence that language-model agents improve democratic decisions.

Question this helps explore

What earlier idea about people and computers does the paper invoke?

Evidence and limits

The authors cite this as background for their interest in human-computer cooperation; their study does not test Licklider’s claim.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 2

What the idea means

Parties and candidates usually offer groups of policies together. The authors suggest that an agent could estimate each person’s preferences and explore systems that separate those policies.

Why it matters

This could reduce the effort of expressing detailed preferences and make some forms of participation easier to organize. The paper does not establish that such systems improve decisions.

Question this helps explore

How might software agents change the choices citizens can express?

Evidence and limits

The authors present this as a way to explore more detailed participation, not as a tested replacement for representative democracy.

Addresses: The proposal for augmented democracy

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 5

What the idea means

The authors say people can interact with language models through ordinary language and that many providers offer them. They see these features as relevant to systems that handle written policies and preferences.

Why it matters

These features may make the technology easier to connect to citizens’ stated views, but they do not establish that it captures those views accurately.

Question this helps explore

Why do the authors consider language models for this task?

Evidence and limits

These are reasons for exploring the technology; they do not show that models are safe or effective representatives.

Addresses: The proposed choice of technology

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 5

What the idea means

The authors cite prior work using language models to simulate poll responses and predict choices in binary elections.

Why it matters

These results help explain why researchers consider models for representing preferences. They do not show that a model can represent every citizen or work reliably in other settings.

Question this helps explore

What earlier evidence does the paper cite about simulated public opinion?

Evidence and limits

These findings are attributed to earlier studies, not to the paper’s own experiment.

Addresses: Prior research on simulated opinions

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 5

What the idea means

The paper mentions prompts that describe personas and additional model training as possible ways to elicit preferences from information in training text.

Why it matters

These methods could support preference estimates, but the passage does not show that those estimates match the views of the people being represented.

Question this helps explore

How might language models produce preference estimates?

Evidence and limits

The authors describe this as a possibility, not proof that models contain an accurate record of people’s views.

Addresses: Possible ways to elicit preferences from models

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 5

What the idea means

The authors say a market with many models may let people leave one provider for another if they fear manipulation or control.

Why it matters

The ability to switch is presented as a possible safeguard, not a guarantee that providers are independent or trustworthy.

Question this helps explore

Why might a range of model providers matter?

Evidence and limits

The paper presents provider choice as a possible protection, while also warning that competition does not remove risks of capture or manipulation.

Addresses: Provider competition as a possible safeguard

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 5, 6

What the idea means

Some models are proprietary and do not openly explain their training data. Some open-source models still depend on private-sector researchers, while others rely on venture funding or government support. The authors say all these arrangements can be vulnerable to capture or manipulation.

Why it matters

Calling a model open-source, for example, does not by itself settle who can influence it or how transparent it is.

Question this helps explore

How might a model’s governance create risks?

Evidence and limits

The authors cite proprietary, open-source, venture-funded, and government-supported models as examples of different arrangements.

Addresses: Ownership and support of language models

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 6

What the idea means

The authors cite work reporting moral judgments similar to people’s and work using detailed personas to predict electoral behaviour.

Why it matters

These results provide background for studying model-based agents, but do not establish that simulated people faithfully represent citizens across settings.

Question this helps explore

What earlier findings about simulated people does the paper report?

Evidence and limits

These findings are attributed to earlier studies, not to the authors’ own experiment.

Addresses: Background research on simulated participants

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 7

What the idea means

In the cited urban-planning work, people compared streetscapes and machine-learning models used their choices to create detailed preference maps. The authors say their study uses a similar setup for policy proposals.

Why it matters

This gives the study a way to learn from limited comparisons, but the passage does not show that the method transfers equally well from streets to policy.

Question this helps explore

What earlier method inspired the study’s design?

Evidence and limits

The authors describe this as an analogy for their approach, not proof that the two applications have identical results.

Addresses: The study’s methodological model

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 7

What the idea means

The authors use people’s policy choices to explore ways of augmenting preference data and to develop tests for individual preferences and group preferences.

Why it matters

These tests address both person-level prediction and whether a sample can reproduce a group pattern. Their value depends on the study’s data and design.

Question this helps explore

What does the study aim to contribute?

Evidence and limits

This is the authors’ stated aim, not a conclusion about how well the methods work.

Addresses: The paper’s stated contribution

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 7

What the idea means

People using Brazucracia.org chose between pairs of proposals drawn from the government programs of the two main presidential candidates. The paper gives examples of proposals shared by both programs and one included in Lula’s program.

Why it matters

These choices let researchers study stated preferences among participants in that experiment. They do not represent all policy choices or all Brazilian voters.

Question this helps explore

What data did the study use?

Evidence and limits

The data come from the named experiment and election period; they should not be treated as a survey of all Brazilians.

Addresses: The study’s participant data

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 8

What the idea means

The experiment collected details including age, political ideology, rural or urban location, education, gender, and geographic location. The authors report research ethics approval and say the data were made public with the earlier publication.

Why it matters

These details help describe the study and its data practices, but do not remove limits in who took part or what the responses can show.

Question this helps explore

What participant information and research procedures does the paper report?

Evidence and limits

These details describe what the authors report; they do not establish that the sample represents the public.

Addresses: Study data and procedures

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 8

What the idea means

Participants made 8,719 choices between pairs drawn from 67 proposals. The paper describes the data as sparse, rather than as a complete record of each participant’s preferences.

Why it matters

The incomplete data motivate methods for estimating preferences, but also limit what can be learned directly from observed choices.

Question this helps explore

How large and complete was the preference dataset?

Evidence and limits

The authors identify the limited set of recorded comparisons as a reason to study augmentation.

Addresses: The study’s data coverage

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 8

What the idea means

The authors report a 95.38% reliability estimate for aggregate preferences in the full participant sample and say it is consistent with prior paired-comparison research, where rankings tended to converge after about 30 to 40 choices per participant.

Why it matters

The comparison offers context for interpreting the reported stability. It does not establish that the estimate is independently validated, that each participant’s preferences were stable, or that the sample represents Brazil.

Question this helps explore

How stable were the group-level preferences, and what context do the authors give for the estimate?

Evidence and limits

The comparison with earlier work provides context for the estimate; it does not independently validate it or show that the participants represent Brazil.

Addresses: Reliability of the full sample’s group preferences

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 8

What the idea means

The study tests whether a smaller sample of these participants can help predict preferences across the full group of 267. It does not establish what people across Brazil prefer.

Why it matters

This scope matters when judging how far the findings can be applied beyond the people who took part.

Question this helps explore

What population do the study’s results describe?

Evidence and limits

The authors define the 267 participants as the group whose preferences they aim to model. They state that the data cannot support estimates for Brazil’s general population.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 9

What the idea means

The test checks whether a model trained on one group can predict choices made by people whose data it did not train on.

Why it matters

Holding out test data gives a check on prediction beyond the examples used for training, though it does not show how a system would perform in real decisions.

Question this helps explore

How do the researchers test whether a model predicts individual choices?

Evidence and limits

The described cross-validation setup separates participants into training and test groups. The model’s predictions are compared with choices in the test group.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 9

What the idea means

This analysis asks how well smaller samples, with or without model-generated additions, match the combined choices of the complete participant group.

Why it matters

It evaluates a different task from predicting one person’s choice. The passage does not provide the results of this comparison.

Question this helps explore

How do the researchers assess prediction of group preferences?

Evidence and limits

The paper describes comparing pure and augmented random samples against the aggregate preferences of all 267 participants.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 9

What the idea means

The models learn from choices people made between proposals, alongside information participants reported about themselves.

Why it matters

The setting and data type help define what the results can speak to; they do not by themselves show how models would perform with other kinds of public input.

Question this helps explore

What data source underlies the model tests?

Evidence and limits

The passage identifies the online project and describes its role in collecting participant choices and demographic data.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 9

What the idea means

The researchers fine-tuned models for this task using examples of participants’ choices. They retrained about 12% of the parameters and left the rest fixed.

Why it matters

These are study methods, not proof that order effects were removed or that the method works equally well with other data.

Question this helps explore

How did the researchers adapt models to predict participants’ choices?

Evidence and limits

The method uses participant information and choice examples in prompts. Reversing proposal order is intended to address sensitivity to which option appears first. LoRA adapts a portion of model parameters without retraining all of them.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 10

What the idea means

The study compares models with different sizes and training backgrounds, including a simpler benchmark. Open-weight means the model weights are available, but it does not necessarily mean the training data were public.

Why it matters

These differences matter when interpreting what the comparison can show. It covers the models the authors could fine-tune at that time, not every available model.

Question this helps explore

Which models were compared, and what distinguishes their training backgrounds?

Evidence and limits

The authors include a simpler benchmark to examine whether a larger model is needed for this task. They also report differences in models’ access to weights and training data. GPT-4 and Claude 3 were not included because fine-tuning options were not publicly available to the authors at the time.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 11

What the idea means

The figures show how often each model matched recorded pairwise choices in this test of the 267 participants; the values after ± are the reported uncertainty amounts.

Why it matters

The results compare models on this dataset and test design. They do not establish prediction accuracy for Brazil’s wider population or for real decisions.

Question this helps explore

How accurately did the models predict individual choices in the reported test?

Evidence and limits

The authors report these accuracy figures for a test set made up of half the participants, after training on the other half. They calculate 95% confidence intervals using 100 bootstrap iterations.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 11, 12

What the idea means

Accuracy counts matching choices. Kendall’s tau compares agreement across pairs. In the authors’ outcome coding, a single choice contributes linked positive and negative results, so the usual precision, recall, and F1 calculations have the same value as accuracy.

Why it matters

This explains why the paper does not report F1 as a separate measure. The passage does not give Kendall’s tau results.

Question this helps explore

How do the authors measure agreement between predicted and recorded choices?

Evidence and limits

The passage describes accuracy and Kendall’s tau and explains how the paired-choice data link the positive and negative outcome counts used to calculate F1.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 12

What the idea means

The planned comparisons cover political ideology, older and younger age groups, education, and sex. Balancing the groups means reducing the larger group to make its size match the smaller group for the comparison.

Why it matters

This method helps compare reported accuracy across the groups studied, but the supplied passages do not give those results or show that the groups represent people beyond the participants.

Question this helps explore

How do the authors plan to examine differences in prediction across participant groups?

Evidence and limits

The passage introduces demographic accuracy comparisons as a way to explore possible model bias. The figure caption names the groups and says the samples were balanced by downsampling the more represented group.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 9, 10, 12

What the idea means

The study includes additional analyses using different shares of the participant group to train the models.

Why it matters

These analyses address how results vary with training-sample size. The supplied passage does not give their outcomes.

Question this helps explore

Did the researchers examine how training-sample size relates to model results?

Evidence and limits

The authors say that results for fine-tuned models trained on random samples of 5%, 25%, and 75% are presented in Supplementary Appendix E.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 12

What the idea means

The authors compared predictions with participants’ reported choices within groups. When one group was larger in the training data, they retrained using a smaller, balanced sample.

Why it matters

This makes group-by-group accuracy comparisons possible, while leaving open whether results would hold with other samples or ways of balancing.

Question this helps explore

How did the authors compare model accuracy across demographic groups?

Evidence and limits

For example, their method matched the number of people from two groups by randomly sampling from the larger group. This helps describe how they checked differences, but does not establish that balancing removes every possible source of bias.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 13

What the idea means

The authors report statistically significant accuracy gaps by political view and education across the six tested models.

Why it matters

The gaps raise a practical question about whose choices these models predict well, but they do not explain why the gaps occurred.

Question this helps explore

Did prediction accuracy differ by political view or education in this study?

Evidence and limits

The reported differences matter because a model that predicts some groups less accurately may represent their choices less well. These results concern the participants and models in this study, not all people or all language models.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 13

What the idea means

Three named models showed a slight, significant tendency to predict younger-quartile participants more accurately. The reported sex comparison favored female participants for three models and male participants for two models.

Why it matters

Accuracy differences may depend on the model and the groups compared, so a result for one model or age comparison should not be assumed to apply more broadly.

Question this helps explore

How did accuracy differences by age and sex vary across models?

Evidence and limits

The age result applies to comparisons between the younger and older quartiles, not to every possible age group. The sex results also differed across models.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 13

What the idea means

The rule links a participant’s reported political view to a candidate’s program. The authors compared choices of proposals from each named program with choices from other programs, separately for liberals and conservatives.

Why it matters

This clarifies the benchmark’s tested comparisons. It does not show that all voters choose according to this rule.

Question this helps explore

How did the authors define and test the party-based comparison rule?

Evidence and limits

The mapping and contrasts clarify what the benchmark tested. The separate treatment of same-program pairs defines what the first comparison covered.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 14

What the idea means

The reported accuracy favored the models in most comparisons with the party-based rule. The authors also report significant predictive power for choices between proposals from the same program, where the rule could not make a prediction.

Why it matters

The results suggest that the models predicted more than this particular orientation-based rule captured in these tests. They do not show that the models would predict choices better in other settings.

Question this helps explore

How did the models compare with the party-based rule and with predictions based only on political orientation?

Evidence and limits

The results favor the models in most reported comparisons, but include specific exceptions. The authors also report significant predictive power for same-program choices. Their overall conclusion is limited to the study’s models, participants, proposals, and comparison methods.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 15

What the idea means

A proposal’s win rate is its number of recorded wins over other proposals divided by the recorded pairwise comparisons involving it. The authors compare these rates for the full sample and a random sample.

Why it matters

The measure lets the authors examine whether a smaller sample can reproduce the study sample’s relative levels of support for proposals.

Question this helps explore

How did the authors measure and compare group-wide policy preferences?

Evidence and limits

This approach gives a way to summarize choices even when participants did not compare every possible pair of proposals. Its conclusions concern how well a sample reproduces the full study sample, not necessarily the preferences of a wider public.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 16

What the idea means

The authors calculate the correlation between win rates from the sample and from all participants, then square it. They describe 100% as identical win rates.

Why it matters

This gives a single score for similarity in proposal-level win rates, but does not measure every kind of representativeness.

Question this helps explore

How did the authors judge whether a sample reproduced the full sample’s preferences?

Evidence and limits

This describes the study’s chosen measure of similarity. A high match on this measure concerns the pattern of proposal win rates; it does not alone establish that every individual’s choices or every other feature of the group is represented.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 16

What the idea means

The authors compare a probabilistic sample on its own with that sample plus an LLM trained on the same sample, using the full set of 267 participants as the reference.

Why it matters

The comparison asks whether the model adds value for estimating the study group’s aggregate preferences. Passage 16 states the question and method, not the result of this comparison.

Question this helps explore

Does adding a model to a sample improve estimates of the full study group’s preferences?

Evidence and limits

This is the study’s stated comparison for estimating group-wide preferences. It is about reproducing the preferences measured among these 267 participants, not necessarily those of a wider public.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 16

What the idea means

The authors say their group-accuracy analysis contributes to discussion about whether language models may overrepresent some segments of the public.

Why it matters

The concern gives a reason to examine accuracy across groups, but the reported accuracy gaps alone do not establish overrepresentation.

Question this helps explore

Why do the authors consider group differences in prediction accuracy important?

Evidence and limits

The passage links the observed accuracy differences to a concern about whose views may be represented well. It does not show that the models actually overrepresent any group.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 13, 14

What the idea means

The authors describe the rule as a way to represent voting among policy bundles associated with parties or politicians.

Why it matters

This framing clarifies what the comparison tests, while leaving open whether it captures how voters choose in practice.

Question this helps explore

Why did the authors use a party-based rule as a benchmark?

Evidence and limits

This explains the benchmark’s framing, not a claim that voters always make choices this way.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 14

What the idea means

The methods differ in whose preferences are predicted and whether the predictions concern existing or new proposals. Type II C combines predictions about existing and new proposals for people already in the sample; Type III also adds new people.

Why it matters

Keeping these methods distinct helps readers understand what information each approach adds and what the reported tests compare.

Question this helps explore

What kinds of preference data can a language model add?

Evidence and limits

The passage distinguishes methods by whether they add people, preferences, or proposals. Type II C combines the two Type II methods; Type III combines predictions for new people and new proposals.

Addresses: The augmentation methods described in the study

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 17, 18

What the idea means

A proposal's win rate is how often participants chose it over another proposal. R² is used to summarize how closely the win rates from a smaller or augmented sample track those from the full participant data.

Why it matters

This tells readers what the reported accuracy gains mean: they concern reproducing the study's recorded pattern of preferences.

Question this helps explore

How did the authors measure whether an augmented sample matched the full participant data?

Evidence and limits

This is a measure of similarity to the full study sample, not a direct test of whether the preferences are correct for people outside the study.

Addresses: The study's measure of sample accuracy

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 17

What the idea means

The authors compared random samples with samples augmented using predictions about additional people. They report the largest stated gain for the 5% sample and a smaller stated gain for the 25% sample; 50% was also tested, but no separate result for it is given in the supplied text.

Why it matters

The stated results suggest that predictions helped reproduce the recorded preferences of the full study sample under these test conditions. They do not establish that the predictions reflect the preferences of people outside the study.

Question this helps explore

Did adding model predictions for more people improve the study's measure of sample accuracy?

Evidence and limits

These figures describe tests on the study's data, with models fine-tuned on the same data and predictions added for a specified share of people outside the initial sample. The text identifies 50% as a tested sample size but gives no separate numerical result for it. These tests do not show how well the method would work in other settings.

Addresses: The reported result for Type I B augmentation

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 18, 19

What the idea means

At zero temperature, the non-fine-tuned models always chose option A, so the authors did not report their results. A majority-vote combination of model predictions performed no better than single models.

Why it matters

These results caution against assuming that a model combination or a model without additional training will improve preference estimates.

Question this helps explore

Did the study's simpler model setups or combined predictions improve the results?

Evidence and limits

The findings concern the specific model settings described; they do not show that all non-fine-tuned models or all ways of combining predictions fail.

Addresses: Reported results for non-fine-tuned models and a majority-vote ensemble

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 19

What the idea means

The authors varied the listed demographic categories in tests with GPT-3.5 Turbo and Mistral 7B. They also compared English and Portuguese prompts and consultation, and report little change in accuracy.

Why it matters

These results suggest that accuracy was not strongly tied to those particular choices. They do not establish equal performance across other groups, languages or settings.

Question this helps explore

How sensitive were the reported results to the tested demographic details and language?

Evidence and limits

The report concerns particular demographic categories, two models and an English-to-Portuguese change. It does not show that the method works equally well for all groups or languages.

Addresses: Reported tests of variation in model inputs

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 18, 19

What the idea means

Type II A and B predict additional preferences for people already represented in the sample. In the reported comparison, they used all demographic information and gave stronger results than Type I B, which predicts preferences for additional people.

Why it matters

The comparison suggests that recorded preferences may help in these tests, but the authors' proposed explanation remains uncertain and the result is limited to the study's data and model.

Question this helps explore

How did Type II augmentation compare with Type I B, and what explanation did the authors offer?

Evidence and limits

The reported comparison is for tests using all demographic information. The explanation about the extra preference information is the authors' conjecture, not a separately established cause.

Addresses: Reported results and proposed explanation for Type II A and B augmentation

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 17, 18, 19

What the idea means

LoRA is a method for fine-tuning a model. RAG adds reference documents when the model responds. In this test, the reference summarized proposals from the participant's favored candidate, with the candidate depending on the participant's political group.

Why it matters

The comparison suggests that the tested fine-tuning method matched these participants' recorded preferences better than this particular reference-document approach.

Question this helps explore

Did fine-tuning or added reference material better match the full participant data?

Evidence and limits

This comparison used one model and reference documents selected according to participants' favored candidates. It does not establish that fine-tuning is always better than supplying context.

Addresses: The study's comparison of fine-tuning with retrieval-augmented generation

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 20

What the idea means

The authors say fine-tuning could make software agents that represent an individual's preferences more accurately than agents given extra context in a chat. The evidence described is about matching preferences in the study data.

Why it matters

The result motivates further testing, but it does not establish how these agents would perform in real civic participation or decision-making.

Question this helps explore

What does the authors' fine-tuning result suggest about personalized AI helpers?

Evidence and limits

A better match on this study's measure supports the authors' proposal for personalized agents, but does not by itself show that such agents work safely or well in real civic decisions.

Addresses: The authors' interpretation and its evidential limit

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 20

What the idea means

Type III augmentation predicts preferences for both additional people and additional proposals. The authors say the Type III A results were similar to the other augmentation results and note that all demographic information was used.

Why it matters

This adds a tested combination of the two kinds of augmentation, but the passage does not provide a distinct numerical estimate for it.

Question this helps explore

What did the authors report for Type III A augmentation?

Evidence and limits

The passage reports a qualitative similarity to other tests, but gives no separate numerical result here. The Type III category combines new people and new proposals.

Addresses: The reported Type III A augmentation result

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 17, 19

What the idea means

A participant's choices between pairs of proposals are represented in a matrix. The authors sample and augment those recorded choices, then assess how well the resulting sample matches the full sample's aggregate preferences.

Why it matters

This clarifies what the reported sample match is based on: recorded choices between proposals, rather than a direct test of people's preferences outside the study.

Question this helps explore

What data did the authors sample and augment in their validation procedure?

Evidence and limits

This describes the data unit and the comparison used to assess whether a smaller or augmented sample reproduces the full participant data.

Addresses: The preference-data structure and validation setup

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 17

What the idea means

The authors say they examined sensitivity to temperature, a model setting, and report those tests in an appendix.

Why it matters

This shows that temperature was checked, but the supplied passage does not reveal whether changing it affected the results.

Question this helps explore

Did the authors check whether changing model temperature affected the results?

Evidence and limits

Passage 19 notes that the tests are presented in Appendix H but gives no outcome, so no conclusion about temperature's effect can be drawn here.

Addresses: The reported temperature sensitivity tests

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 19

What the idea means

The cited study suggests that many participants were open to AI sharing decision-making power, while politicians retained most of it.

Why it matters

Public views may allow discussion of limited AI roles, but the result does not show that people would accept such a system in practice.

Question this helps explore

Would people accept AI having some influence on public decisions?

Evidence and limits

This is a result reported from another study, not a finding from the authors’ own experiment. The passage does not describe the study’s sample in detail.

Addresses: Public acceptance of AI influence

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 21

What the idea means

Participants showed some openness to AI having influence on technical matters, but still preferred experts.

Why it matters

Openness to AI influence can coexist with a preference for expert judgment. The passage does not say that this preference depends on whether experts are available.

Question this helps explore

Do people want AI to make technical decisions?

Evidence and limits

This is a finding from the study cited by the authors. The passage does not specify which technical tasks were considered.

Addresses: Public views on AI in technical decisions

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 21

What the idea means

The authors describe several AI-candidate efforts as receiving more media attention than electoral success. The Tama City example reached the second round in 2018, but its vote total was lower in 2022.

Why it matters

Interest in AI candidates, or approval in a survey, does not by itself show that voters will choose them.

Question this helps explore

Have AI candidates gained electoral support in practice?

Evidence and limits

The authors use these examples to distinguish stated willingness to consider AI from voting for an AI candidate. The passage does not explain why support changed.

Addresses: Electoral results for centralized AI candidates

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 21

What the idea means

The researchers adjusted six available language models using detailed political preference data collected during the Brazilian election.

Why it matters

The results concern these models, this data, and the study tasks; they do not establish how a full civic system would work.

Question this helps explore

What data and models did the authors study?

Evidence and limits

The paper describes using organically collected, fine-grained preference data to study model predictions. This was not a test of a complete voting system.

Addresses: Study method and data

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 22

What the idea means

The models did not perform equally across the participant groups examined: prediction was better for people reporting these characteristics.

Why it matters

Uneven accuracy matters if an AI helper is meant to represent people fairly, because some people’s preferences may be reflected less well.

Question this helps explore

Did prediction accuracy differ between groups of participants?

Evidence and limits

This is a result from the authors’ study. It describes group-level differences and does not mean every person in those groups was predicted more accurately.

Addresses: Differences in prediction accuracy

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 22

What the idea means

The researchers used a limited sample to predict the overall preferences of all participants, then compared those estimates with estimates from random samples that were not augmented.

Why it matters

If the result holds elsewhere, this approach could help estimate group preferences when few people respond. It needs testing beyond this study.

Question this helps explore

Can AI help estimate a group’s preferences from a small sample?

Evidence and limits

The authors compare estimates of the full participant group’s preferences. They do not describe filling in individual participants’ missing responses.

Addresses: Model-based sample augmentation

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 22

What the idea means

In the study comparison, model predictions were more accurate than predictions based only on a person choosing the policies linked to their preferred party or politician.

Why it matters

People’s policy preferences may be more varied than a simple party label captures, but the comparison does not reveal how the models made their predictions.

Question this helps explore

Did the models do better than a simple party-based prediction rule?

Evidence and limits

The authors interpret the comparison as evidence that their models captured preference details beyond a simple left-right political divide. It does not show what specific information the models used.

Addresses: Comparison with party-based prediction

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 22

What the idea means

The authors see potential for research but judge that the systems they discuss are not ready to be used as full civic decision-making systems.

Why it matters

Study results should not be mistaken for evidence that these tools are safe or reliable enough for real public decisions.

Question this helps explore

Are these systems ready to use in real civic decisions?

Evidence and limits

This is the authors’ overall assessment of the technology discussed in the paper, not a claim that every possible future system will remain unsuitable.

Addresses: Readiness for real-world use

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 24

What the idea means

When people cannot inspect a model’s training data, they may have limited ability to assess or challenge how it was shaped. The authors suggest that open code and data rules deserve attention.

Why it matters

Transparency could help people examine systems used in sensitive public decisions, but the passage does not specify a complete set of safeguards.

Question this helps explore

What transparency concerns do the authors raise?

Evidence and limits

The paper raises this as a concern and a direction for developing systems; it does not establish that open code or data alone would prevent manipulation.

Addresses: Transparency in system development

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 24

What the idea means

A model might favor one option when it appears first, but give a different result when the same options are presented in reverse order.

Why it matters

If results depend on presentation order, people may have difficulty trusting or checking the model’s predictions.

Question this helps explore

Do the models give consistent answers when choices are presented differently?

Evidence and limits

The authors report this result for some cases and point to a supplementary section; the passage does not say that all models or predictions had this level of inconsistency.

Addresses: Consistency of predictions

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 24

What the idea means

The authors found differences in how well models predicted people’s preferences and report that some models showed a tendency to favor female participants.

Why it matters

A system that represents citizens should be checked for differences in how it treats groups, not only for its overall accuracy.

Question this helps explore

What bias did the authors observe in their own study?

Evidence and limits

The reported results concern the models and participants examined in this paper. The passage does not provide the size of these effects.

Addresses: Group differences and gender bias

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 24

What the idea means

The authors point to a tension: people may worry about technology’s effects on society while some are still willing to give AI a role in civic decisions.

Why it matters

Public concern alone does not show that people reject every civic use of AI. The passage presents this as framing, not as a result measured in the authors’ study.

Question this helps explore

Can concern about digital technology coexist with openness to AI in public decisions?

Evidence and limits

This is the authors’ framing in the discussion, not a finding from their own experiment. The passage does not quantify either attitude.

Addresses: Public attitudes toward digital technology and AI in civic decisions

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 21

What the idea means

The authors say they do not have a thorough account of why the models select one proposal rather than another.

Why it matters

Without a clear explanation, people may find it harder to check or question model predictions used to represent political preferences.

Question this helps explore

Can people understand why a model favors one proposal?

Evidence and limits

The authors identify this as a limitation of their approach. The passage does not propose or test a way to address it.

Addresses: Explaining model predictions

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 24

What the idea means

The reported study examined whether models could reproduce survey responses from different groups. Its gender comparison found only a small accuracy difference.

Why it matters

These results suggest that prediction accuracy can vary between groups, but the reported difference was small and does not establish how well the method works in other settings.

Question this helps explore

How well have language models reproduced the views of different groups?

Evidence and limits

This is a finding attributed to an earlier study, not a new result from the publication being discussed.

Addresses: Reported findings from Argyle and colleagues

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 25

What the idea means

The reported study found stronger reproduction of preferences for several demographic groups and a small accuracy difference by gender.

Why it matters

A model’s overall accuracy may hide differences between groups, so results for one group should not be assumed to apply equally to everyone.

Question this helps explore

Which groups did the reported models predict more accurately?

Evidence and limits

These are findings attributed to an earlier study; the passage reports them without explaining the study’s methods.

Addresses: Reported findings from Santurkar and colleagues

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 25

What the idea means

The authors say they do not fully understand what leads a model to choose one proposal over another.

Why it matters

Without that understanding, people may find it difficult to check or question the model’s role in civic decisions.

Question this helps explore

Can people understand why a model favors one proposal?

Evidence and limits

The passage identifies the lack of understanding of model choices as a limitation of the proposed approach.

Addresses: The authors’ stated limitation

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 25

What the idea means

The authors propose examining data augmentation with data gathered in other ways, such as through voting advice applications.

Why it matters

This would extend the kinds of data-collection cases studied. The passage does not state whether using these sources would improve results.

Question this helps explore

What other kinds of collected data could be explored?

Evidence and limits

The authors identify these data sources as an avenue for further study, without stating that they should improve the approach.

Addresses: A proposed direction for future research

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 25

What the idea means

The proposal is to guide models toward different political views and have other models summarize those views.

Why it matters

This could be studied as a way to present differing views, but the passage does not show that the summaries would be accurate or fair.

Question this helps explore

Could models represent and summarize different political viewpoints?

Evidence and limits

The passage presents this as an avenue to explore and refers to earlier work; it does not report a finding from the proposed test.

Addresses: A proposed direction for future research

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 25

What the idea means

The authors point to a proposal that combines external information with further training of a model.

Why it matters

Testing the combination could show whether it improves the approach, but the passage reports no such result.

Question this helps explore

Could adding information and further training improve data augmentation?

Evidence and limits

The passage identifies this as a possibility to explore, not a demonstrated improvement.

Addresses: A proposed direction for future research

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 25

What the idea means

The authors see AI’s potential role in supporting democracy as an area that has not been thoroughly explored and call for research involving different fields.

Why it matters

Their proposals and findings are part of an area they say needs further exploration.

Question this helps explore

What do the authors say is needed to explore AI’s role in democracy?

Evidence and limits

The passage calls the area uncharted and says that further advances will require interdisciplinary research.

Addresses: The authors’ view of the research area

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 25, 26

What the idea means

The limitations discussed in the paper do not cover every possible problem with the approach.

Why it matters

Readers should not treat the listed concerns as a complete account of what would need attention.

Question this helps explore

Are the limitations described in the paper a complete list?

Evidence and limits

The authors explicitly describe the limitations as non-exhaustive.

Addresses: The authors’ qualification about the approach’s limitations

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 25

What the idea means

The authors present their findings as a possible starting point for more organized research, not as a settled account of how AI should support democracy.

Why it matters

This clarifies the intended role of the paper: to encourage further study in an area the authors consider largely unexplored.

Question this helps explore

What contribution do the authors hope their findings will make?

Evidence and limits

The authors describe their hoped-for contribution as helping stimulate and organize exploration of this largely unexplored area.

Addresses: The authors’ stated aim for the findings

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 26

What the idea means

The researchers worked with a curated set of proposals attributed to Luiz Inácio Lula da Silva, Ciro Gomes, and Jair Bolsonaro. The lawyer’s role was part of the curation process.

Why it matters

The experiment’s results depend in part on which proposals were selected and how they were described.

Question this helps explore

Where did the study’s policy proposals come from?

Evidence and limits

This describes the source and curation of the proposals, not a check that the proposals were complete or fair representations of the candidates’ platforms.

Addresses: Study materials

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 34

What the idea means

The authors gave two reasons for using small sets: ordering preferences across all 67 proposals was challenging, and earlier research reported difficulty predicting choices among three or more proposals.

Why it matters

The experiment focused on pairs, so its results may not show how well models handle larger sets or the full list of proposals.

Question this helps explore

Why did the experiment focus on pairs of proposals?

Evidence and limits

The difficulty ordering all 67 proposals is a reason stated by the authors. The claim about model difficulty with larger sets comes from earlier research, not a new result from this study.

Addresses: Experimental design

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 34

What the idea means

The authors report that they did not run an original study with human participants or gather new primary data.

Why it matters

The passages do not provide direct evidence here about how actual voters would respond to the proposals or the system.

Question this helps explore

Did this study collect new data directly from people?

Evidence and limits

This is a statement about the study’s methods, not proof that the proposals or any model outputs reflect what voters think.

Addresses: Study limits

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 34

What the idea means

The proposal concerns access to banking and tax-secrecy information for officials at two levels of the executive branch.

Why it matters

It raises a transparency option for comparison, but the passage does not explain its legal design or limits.

Question this helps explore

What transparency proposals appear in the comparison set?

Evidence and limits

The table attributes this proposal to Ciro; the passage does not define the covered information or safeguards.

Addresses: Table 1 policy proposals

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 37

What the idea means

The proposal concerns transparency about the ultimate recipients or beneficiaries of public resources.

Why it matters

It is a transparency option in the comparison set, but its scope and safeguards are not explained.

Question this helps explore

What transparency proposals appear in the comparison set?

Evidence and limits

The table attributes this proposal to Bolsonaro; it does not define the disclosure rules or their reach.

Addresses: Table 1 policy proposals

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 37

What the idea means

The proposal connects greater transparency with compliance with the Access to Information Law.

Why it matters

It is a listed route to transparency, but the excerpt does not describe enforcement or implementation.

Question this helps explore

What transparency proposals appear in the comparison set?

Evidence and limits

The table attributes this proposal to Lula; it does not specify which changes would improve compliance.

Addresses: Table 1 policy proposals

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 37

What the idea means

The proposal links better and less expensive transport services to the structuring of concessions and public-private partnerships.

Why it matters

It presents a proposed means of improving services, but does not show that these arrangements would lower prices.

Question this helps explore

What proposals address transport services and prices?

Evidence and limits

The table attributes this proposal to Bolsonaro; it does not describe the concession or partnership terms.

Addresses: Table 1 policy proposals

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 37

What the idea means

The proposal calls for encouraging agricultural practices described as sustainable.

Why it matters

It is a policy option in the comparison set, but the passage does not specify the practices or how they would be encouraged.

Question this helps explore

What proposals address agricultural practices?

Evidence and limits

The table attributes this proposal to Bolsonaro and Lula; it does not define which practices count as sustainable.

Addresses: Table 1 policy proposals

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 37

What the idea means

The proposal is to support agro-industry and produce inputs within the country.

Why it matters

It is a listed policy option, but the passage does not specify which industries or inputs are meant.

Question this helps explore

What proposals address agro-industry and domestic inputs?

Evidence and limits

The table attributes this proposal to Bolsonaro and Lula; the entry continues across the page break.

Addresses: Table 1 policy proposals

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 37, 38

What the idea means

The proposal links expanding clean and renewable energy sources with strengthening the energy supply.

Why it matters

It is a listed policy goal, but the excerpt does not assess reliability, cost, or environmental effects.

Question this helps explore

What proposals address energy supply?

Evidence and limits

The table attributes this proposal to Bolsonaro and Lula; it does not specify technologies or targets.

Addresses: Table 1 policy proposals

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 38

What the idea means

The proposal is to offer green bonds to finance investments considered sustainable, with transport and energy named as areas.

Why it matters

It introduces a possible funding approach, but the passage does not explain bond terms or how sustainability would be judged.

Question this helps explore

What proposals address funding for sustainable investments?

Evidence and limits

The table lists transport and energy as examples and says there may be other areas; it does not define how projects qualify.

Addresses: Table 1 policy proposals

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 38

What the idea means

The proposal calls for meeting the country’s carbon-gas reduction targets associated with the 2015 Paris conference.

Why it matters

It records a climate policy goal, but does not state how the targets would be met.

Question this helps explore

What climate-related proposals appear in the comparison set?

Evidence and limits

The table attributes this proposal to Lula; it does not specify the actions or measures of progress.

Addresses: Table 1 policy proposals

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 38

What the idea means

The proposal calls for recovering land degraded by destructive activity and reforesting areas that have been devastated.

Why it matters

It is a listed restoration aim, but the passage does not show what land or resources would be involved.

Question this helps explore

What proposals address damaged land and forests?

Evidence and limits

The table attributes this proposal to Bolsonaro and Lula; it does not specify locations, scale, or methods.

Addresses: Table 1 policy proposals

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 38

What the idea means

The listed proposal connects more ecological bases with efforts to curb “drug mining” and money laundering in the Amazon.

Why it matters

The wording identifies an enforcement approach, but the unclear phrase and lack of detail limit what can be inferred.

Question this helps explore

What proposals address illicit activity and enforcement in the Amazon?

Evidence and limits

The phrase “drug mining” is unclear in the excerpt, so this summary retains it rather than guessing its intended meaning. The table attributes the proposal to Bolsonaro.

Addresses: Table 1 policy proposals

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 38

What the idea means

Participants saw pairs of proposals and made comparisons through the platform without identifying themselves.

Why it matters

This describes how preferences were collected, but does not by itself show how well the sample represented the wider public.

Question this helps explore

How did the study present policy choices to participants?

Evidence and limits

The passage describes a classical pairwise comparison screen and says participation was anonymous; it does not explain recruitment or voting rules here.

Addresses: Brazucracia comparison platform

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 38

What the idea means

The study’s table lists six kinds of self-reported participant information that could be added to prompts. Age groups use the top and bottom quarters of the sample.

Why it matters

These categories describe the context available for the prompts; they do not show that the categories capture every relevant difference between people.

Question this helps explore

What participant information did the researchers use as context?

Evidence and limits

The table gives age as young or old using the top and bottom quartiles; ideology as conservative, centrist, or liberal; zone as urban or rural; education as college or non-college; sex as male or female; and location as city and state.

Addresses: Table 2 demographic characteristics

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 38

What the idea means

The researchers split participants into two random groups of equal size and compared how strongly the groups’ proposal win rates tracked each other. R² summarizes the strength of that relationship.

Why it matters

This checks whether preferences aggregated across groups look similar as sample size changes; it does not establish that the participants represent the whole population.

Question this helps explore

How did the researchers check whether they had enough participant data?

Evidence and limits

They used the coefficient of determination, R², for the Pearson correlation between each proposal’s win rates in the two samples, and plotted the result against sample size.

Addresses: Appendix B data adequacy checks

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 38, 39

What the idea means

The researchers estimated a 0.72 match measure for two groups of 53 participants and reported 0.9538 for the full sample. The figure caption gives the 53-person result as 72.2%.

Why it matters

These figures suggest that the sample’s aggregated preferences were fairly consistent under this check, but do not show that the preferences match those of the wider population.

Question this helps explore

What did the study report about the stability of participant preferences?

Evidence and limits

The appendix presents the 53-person estimate and full-sample test-retest consistency as results of its data-adequacy analysis, not as proof of population representativeness.

Addresses: Appendix B data adequacy checks

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 39

What the idea means

For each model, the researchers kept its architecture unchanged when comparing different amounts of training data.

Why it matters

This reduces one source of difference between those comparisons, but does not show that all other conditions were identical.

Question this helps explore

How did the researchers try to make model comparisons comparable?

Evidence and limits

The passage says the architecture for each language model was kept the same across training-set sizes; it does not say this controls every other difference.

Addresses: Appendix C model-training methods

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 39

What the idea means

The researchers translated prompts into Brazilian Portuguese with a prompt from Appendix C so they could use BERTimbau embeddings. They state that embedding quality matters greatly for BERT’s predictive accuracy and contextual understanding.

Why it matters

This explains a language-preparation choice and the researchers’ reason for it. The passage does not compare these embeddings with alternatives or show that this choice improved results.

Question this helps explore

How did the researchers prepare prompts and embeddings for the BERT model?

Evidence and limits

The appendix reports the translation procedure, the researchers’ rationale for using embeddings, and the stated source and size of those embeddings.

Addresses: Appendix C BERT methods

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 39, 40

What the idea means

At the time described in the appendix, the authors said they planned to post code for reproducing results, and that the fine-tuned models were publicly available on GitHub.

Why it matters

This gives readers information about possible access to the study materials. The passage does not confirm whether the planned code posting happened later.

Question this helps explore

Could readers access the code and models used in the study?

Evidence and limits

The passage distinguishes the planned availability of the code from the reported public availability of the fine-tuned models.

Addresses: Appendix C code and model repositories

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 39

What the idea means

The authors report that readers could obtain the fine-tuned models, but describe the code as something they planned to publish.

Why it matters

Access to models may help readers inspect the work; the passage does not confirm that the promised code was later posted.

Question this helps explore

Could readers access the models and reproduction code?

Evidence and limits

The passage distinguishes models said to be available from code whose repository link was still to be updated.

Addresses: Availability of study materials

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 41

What the idea means

The researchers adapted their prompt format to Brazilian Portuguese and included participant characteristics and a choice between two proposals. The listed age categories were young and older, described as “third age.”

Why it matters

The passage documents the prompt design, but does not show whether predictions were robust across languages.

Question this helps explore

What participant details and policy choice appeared in the Brazilian Portuguese prompt example?

Evidence and limits

Passages 41–42 describe the Portuguese prompts, list their demographic features and values, and give an example pairing participant details with two policy proposals. They do not report whether this adaptation improved predictions.

Addresses: Portuguese prompt design

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 41, 42

What the idea means

The researchers checked accuracy using random samples that included different shares of participants.

Why it matters

This may help readers judge how accuracy varies with sample size, but no conclusion can be drawn from this excerpt because it does not show the results.

Question this helps explore

Did the researchers check accuracy with random samples of different sizes?

Evidence and limits

Passage 42 says the researchers repeated an accuracy comparison using random samples of three sizes and identifies five models. The figure's results are not available in the supplied text.

Addresses: Accuracy checks across sample sizes

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 42

What the idea means

Kendall rank correlation measures how similarly two sets of ranked choices are ordered.

Why it matters

This can show whether model rankings align, but the excerpt does not allow readers to judge how strong any alignment was.

Question this helps explore

Did the researchers compare the models' preference rankings?

Evidence and limits

The appendix identifies the measure, models, and test-set size, but the figure's values are not supplied.

Addresses: Kendall rank-correlation analysis

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 42

What the idea means

A model counted as consistent when it chose the same proposal whether that proposal appeared first or second. The researchers compared this measure across models and sample sizes.

Why it matters

Accuracy alone may not reveal whether a model's choice depends on how options are presented. The reported excerpt does not give exact rates for each model or sample size.

Question this helps explore

Did models keep their choices when the options changed order, and how was this checked across models and samples?

Evidence and limits

Passages 42–43 define the measure and report that results were presented across models and sizes of probabilistic samples. They give a broad comparison but not exact rates for each model.

Addresses: Consistency under option-order changes

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 42, 43

What the idea means

The researchers examined whether changing the setting that affects prediction variation influenced outputs from these two models.

Why it matters

This could help readers understand how sensitive predictions are to that setting, but no conclusion is possible from the supplied text.

Question this helps explore

Did the researchers check whether temperature affected predictions?

Evidence and limits

The appendix identifies two models tested, but only the figure caption is included and no results are described.

Addresses: Temperature sensitivity checks

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 43

What the idea means

A two-sample t-test checks whether the average accuracy of two groups differs; this version does not require their variances to be equal.

Why it matters

The method helps assess whether group accuracy differences are unlikely to be due to sampling variation, but it does not show whether the model is fair or why groups differ.

Question this helps explore

How did the researchers test for accuracy differences between groups?

Evidence and limits

The passage describes the test, its equal-means starting assumption, and the model training scope shown in the table caption. A small p-value indicates evidence against equal average accuracy, but not the size or direction of a difference.

Addresses: Statistical tests of group accuracy

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 43, 44

What the idea means

The reported tests found evidence that average prediction accuracy was not the same across each listed pair of groups. In the table, *** denotes the 1% significance level, ** the 2% level, and * the 5% level; these comparisons are marked ***.

Why it matters

The results raise a concern that accuracy may vary across participant groups, but they do not establish the direction or practical size of the differences.

Question this helps explore

Which education and political-ideology groups had different average accuracy?

Evidence and limits

Passage 44 reports significant test results for each of six models in these comparisons. It does not report which group had higher accuracy or how large the differences were.

Addresses: Education and ideology comparisons

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 43, 44

What the idea means

The evidence for differences between younger and older groups varied by model. The table defines ***, **, and * as significance at the 1%, 2%, and 5% levels, respectively; the three reported age differences are marked ***.

Why it matters

This variation means the reported age pattern cannot be treated as the same across all tested models. A result without a significance marker does not prove that the groups had equal accuracy.

Question this helps explore

Did age groups differ in accuracy for every model?

Evidence and limits

Passage 44 reports which models’ tests were marked significant and gives the table’s significance thresholds. It does not say which age group had higher average accuracy or give the size of any difference.

Addresses: Younger and older participant comparisons

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 44

What the idea means

The study reports evidence of different average accuracy for five models, while the GPT-3.5 Turbo comparison did not reach the table’s stated significance levels. The table defines ***, **, and * as significance at the 1%, 2%, and 5% levels, respectively.

Why it matters

The pattern suggests results can vary by model, but it does not show which group benefited or whether the differences were practically large. No significance marker does not establish equal accuracy.

Question this helps explore

Did female and male participants have different accuracy across the models?

Evidence and limits

Passage 44 reports different significance outcomes across models and defines the table’s significance thresholds. It does not say which group had higher average accuracy or give the size of any difference.

Addresses: Female and male participant comparisons

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 44

What the idea means

The paper focuses on using language models to build agents that could support civic participation. It does not assess the broad political effects of the internet or other communication technologies.

Why it matters

The prediction tests cannot answer wider questions about how technology affects democracy.

Question this helps explore

What wider question does the paper set aside?

Evidence and limits

The authors place their work in the history of communication technologies but distinguish their specific aim from studying those technologies’ wider effects.

Addresses: The paper’s study of language-model agents

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 4

What the idea means

The authors warn that model limits may affect representation of people with different political views. They name remote and offline Indigenous communities as examples of groups that may be absent from training data.

Why it matters

A system could represent some citizens less well than others. The authors say this risk must be taken seriously before real-world use.

Question this helps explore

Which citizens might a model represent poorly?

Evidence and limits

This qualification limits the idea that models can represent a wide range of people based on text alone.

Addresses: The proposed use of language models as representatives

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 6

What the idea means

The authors describe weaknesses in certain logical tasks and warn that they could be a problem when a decision requires those abilities.

Why it matters

A system’s required skills need to be considered before it is used in real decisions.

Question this helps explore

What kinds of tasks may exceed current model abilities?

Evidence and limits

This is a task-specific caution, not a claim that models fail at every policy task.

Addresses: Limits on using models in decision systems

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 6

What the idea means

The researchers used a fixed setting for reproducibility and report that varying it did not change their results.

Why it matters

This is a reported robustness check for one training setting, not evidence that all modeling choices leave results unchanged.

Question this helps explore

Did a training setting appear to change the reported results?

Evidence and limits

The authors report that other temperature settings did not affect results and cite supplementary material for details.

Addresses: Training the models

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 10

What the idea means

The authors use repeated resampling to calculate intervals around the reported accuracy figures, with different interval levels for the two sets of figures.

Why it matters

These intervals describe uncertainty under the reported calculation method. They do not show how well the results apply beyond the studied participants.

Question this helps explore

How do the authors describe uncertainty around the reported accuracy figures?

Evidence and limits

The passages specify different confidence-interval levels for the individual-model and demographic figures. Bootstrapping is a way of repeatedly resampling data to estimate uncertainty.

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 10, 11

What the idea means

The paper tested how to combine or estimate preferences, not how groups might negotiate agreement with AI agents involved.

Why it matters

Results about preference estimates should not be treated as evidence that these systems can help groups reach acceptable decisions.

Question this helps explore

What forms of group decision-making did the study leave untested?

Evidence and limits

The authors identify these as open research questions and distinguish them from the preference-aggregation task they studied.

Addresses: Limits on the study’s decision-making scope

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 22

What the idea means

Brazucracia asks users to rank relevant policy proposals, including when their preferences may not align perfectly with one party. Other voting advice tools may instead recommend a party or politician.

Why it matters

Results from a platform that ranks proposals may not transfer to tools that recommend political candidates or parties. The study does not establish whether this design difference changes prediction accuracy.

Question this helps explore

Could the way a platform collects preferences change how well AI predicts them?

Evidence and limits

The authors identify the platform difference as a reason to study whether results depend on design. They do not test the effect by comparing platforms.

Addresses: Effect of platform design

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 23

What the idea means

The reported results describe single-agent models, not systems where several agents work together. The authors note that some outside studies report better accuracy from multi-agent systems.

Why it matters

The paper’s results cannot show whether using multiple agents would improve preference prediction in this setting.

Question this helps explore

Could multiple AI agents perform differently from one model?

Evidence and limits

The study itself did not test multi-agent systems; the comparison comes from other research cited by the authors.

Addresses: Single-model study limitation

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 23

What the idea means

The researchers tested a limited set of models and prompt designs, rather than comparing a broad range of current options.

Why it matters

The results may change with other models or prompts, so they do not settle which approach works best.

Question this helps explore

How complete was the model and prompt testing?

Evidence and limits

These are stated limits of the authors’ exploration; the passage does not claim that newer models or different prompts would perform better.

Addresses: Limits of model and prompt coverage

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 23

What the idea means

The models used a small set of details about each participant rather than a rich record of their life or views.

Why it matters

The findings describe predictions from this limited representation, not what a more detailed personal AI helper could do.

Question this helps explore

How much information did the study use to represent each person?

Evidence and limits

The authors contrast their minimal representation with a more detailed example, but do not test whether a richer representation would improve this study’s predictions.

Addresses: Scope of the personal representation

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 23

What the idea means

Some people may have more text, video, or other information available for building detailed AI representations than others.

Why it matters

If richer data improves representations, unequal access could lead to some people being represented better than others. The paper does not test this effect.

Question this helps explore

Could unequal access to personal data affect political representation?

Evidence and limits

The authors identify uneven data availability as a challenge; they do not measure its effects on representation in this study.

Addresses: Unequal availability of data

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 23

What the idea means

A model’s behavior in a language task does not by itself show how it will behave when predicting people’s political preferences.

Why it matters

Bias needs to be checked in the specific task and setting where a model might be used.

Question this helps explore

Can findings about language bias be applied directly to political preference prediction?

Evidence and limits

The authors make this distinction while comparing their work with earlier research. The passage does not deny that language-model bias may affect civic predictions; it warns against assuming the same pattern without evidence.

Addresses: Limits on transferring bias findings across tasks

Large Language Models (LLMs) as Agents for Augmented Democracy · passages 24