Complete underlying reasoning (70 claims and questions)
The main steps above are a reading path. These are all assessed contributions to this argument, including challenges and questions.
Open · Which governing decisions may AI make under a mandate approved by citizens?
Governing authority bears on whether agents can represent preferences while remaining accountable.
Which governing decisions should AI be allowed to make under a mandate approved by citizens, and which must be referred back to citizens or an independent authority?
It is not yet clear which tasks people, rather than AI, would still do when carrying out the rules.
Inspect the contribution and sourcesAnswered under conditions · How might language models produce preference estimates?
The question bears on the proposal for personal agents: it concerns how preference estimates might be produced, while the supplied evidence explicitly leaves their accuracy and civic representativeness unestablished.
How might language models produce preference estimates?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered under conditions · How might software programs support detailed participation?
The question concerns the proposed participation role of personal agents; the supplied claim does not demonstrate an accountable civic system.
How might software programs that act for people support detailed participation?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesPartly answered · Could unequal access to personal data affect political representation?
Unequal data access is an explicit concern in the narrative about personal agents and accountable representation; here it remains a concern rather than a measured effect.
Could unequal access to personal data affect political representation?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · What role do the authors propose for voter-focused AI helpers?
It asks about the proposed role of personalized voter agents, not proof they are accountable civic representatives.
What role do the authors propose for voter-focused AI helpers?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Why might a range of model providers matter?
Provider choice bears on model control and accountability, while remaining only a possible safeguard.
Why might a range of model providers matter?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Have AI candidates gained electoral support in practice?
The results do not establish that personalized agents have been tested as civic systems.
Have AI candidates gained electoral support in practice?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Are the limitations described in the paper a complete list?
The caveat limits the scope of concerns relevant to personalized agents.
Are the limitations described in the paper a complete list?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · What does the authors' result from adjusting a model suggest about AI helpers tailored to individuals?
The question concerns what prediction results imply for personalized civic agents; the narrative addresses their potential role and the limits of preference matching as representation.
What does the authors' result from adjusting a model suggest about AI helpers tailored to individuals?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · How might software programs that act for people change the choices citizens can express?
The question concerns what citizens can express through proposed software agents; the narrative addresses their potential role and the limits of preference matching as representation.
How might software programs that act for people change the choices citizens can express?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Could readers access the code and models used in the study?
Code and model availability is relevant to transparency and control of proposed agents, but the supplied access statement alone does not establish adequate access or safeguards for accountability.
Could readers access the code and models used in the study?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Can people understand why a model favors one proposal?
Unexplained proposal choices bear on whether people can understand and hold personal agents accountable; the evidence identifies an unresolved limitation.
Can people understand why a model favors one proposal?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Why might candidate policy packages fail to match every person’s views?
The imperfect fit of bundled policies is part of the motivation for individualized agents, while preference matching does not itself establish accountable representation.
Why might candidate policy packages fail to match every person’s views?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Do the models give consistent answers when choices are presented differently?
Unstable choices under reversed presentation qualify confidence that an agent reliably reflects an individual’s stated preferences.
Do the models give consistent answers when choices are presented differently?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesPartly answered · Could models represent and summarize different political viewpoints?
The tentative proposal to represent and summarize viewpoints is relevant to personal-agent proposals and the narrative’s concern that their representational ability remains untested.
Could models represent and summarize different political viewpoints?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · How did the models compare with the political-alignment rule for individual choices?
Prediction performance is relevant to personal-agent proposals but does not establish accountable representation in civic decisions.
How did the models compare with the political-alignment rule for individual choices?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · How might a model’s governance create risks?
Model ownership and capture are relevant governance concerns for accountable personal agents.
How might a model’s governance create risks?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · What participant information did the researchers use as context?
The type and amount of personal information used bears on claims that personalized agents reflect an individual’s preferences, without itself establishing accountability.
What participant information did the researchers use as context?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Are these systems ready to use in real civic decisions?
This asks whether predictive-agent proposals are ready for civic use; the answer distinguishes model research from a tested civic system.
Are these systems ready to use in real civic decisions?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · How much information did the study use to represent each person?
The limited representation bears on whether personalized agents can reflect people’s views; it does not establish fuller representation or accountability of a civic agent.
How much information did the study use to represent each person?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesImportant limit · Platform design may matter
The platform’s method for collecting and presenting preferences may shape the data on which personalized agents are evaluated; this remains an untested possibility.
The authors say model performance may depend on platform design. Brazucracia, the platform used in their study, ranks relevant proposals for people whose preferences may not fit one party perfectly; unlike some voting advice tools, it does not recommend a party or politician.
Source coverage: Full readable text of the 37-page research draft, including references and appendices. Figures are not independently interpreted. 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
Inspect the contribution and sourcesObjection · Unequal data access is a concern
Unequal data access may make personalized agents unequally capable of representing people; this is a stated concern, not a measured result.
The authors say richer digital representations could use more or better data, but access to such data is uneven among people.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. 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
Inspect the contribution and sourcesStarting point · Personalized voter helpers
Introduces personalized agents as possible voter intermediaries, while distinguishing that proposal from tested representation in real decisions.
The authors explore personalized software agents that could assist or act as intermediaries for individual voters.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. The paper presents personalized AI helpers as a possible use of language models. Its tests concern predictions from collected data, not whether an AI helper represents a voter in real decisions.
Inspect the contribution and sourcesImportant limit · People were represented with limited data
A short set of demographics and pairwise preferences is narrower than a demonstrated account of a person’s views, authority, or interests.
The study represented each person using a short set of personal background details and preferences between pairs of options. The authors say this is a limited form of a digital twin, meaning a digital representation of a person.
Source coverage: Full readable text of the 37-page research draft, including references and appendices. Figures are not independently interpreted. The authors contrast their limited 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
Inspect the contribution and sourcesImportant limit · Models in the comparison
The distinction between open or proprietary weights and training data is relevant to model control and transparency in personal agents; it does not establish the governance or accountability of a proposed civic agent.
The comparison includes six language models and the BERT SQuAD question-answering test, a simpler benchmark (a standard test for comparing models). The models differ in size and in what is known about their training: GPT 3.5 Turbo has an architecture and data that are not publicly available; LLaMA-2, LLaMA-3, Mistral, and Gemma are open-weight models, meaning their model weights are available, trained on data that are not publicly available; Falcon is open-weight and trained on public data; and BERT is a foundation model, a model trained on broad data that can be adapted for different tasks, trained on public data.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. 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 options to train them further for a particular task were not publicly available to the authors at the time.
Inspect the contribution and sourcesStarting point · Software programs could ease detailed participation
The proposed agents could support individual policy preferences; reduced participation effort is described as an opportunity to explore, not a tested outcome.
The authors propose software programs that can act for people as another way to reduce the effort of participation while allowing people to state preferences on separate policies.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. The paper presents this as an opportunity to explore, not as a tested solution to the effort involved.
Inspect the contribution and sourcesStarting point · Participation between two forms
Individual policy preferences with software agents are the proposed participation mechanism, not a tested accountable civic system.
The authors describe augmented democracy as participation between choosing representatives and deciding directly on policies.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. The approach lets people indicate preferences on individual policy proposals, with software agents proposed as a way to support detailed participation.
Inspect the contribution and sourcesImportant limit · The study develops tests for individual and group preferences
The stated aims do not demonstrate that agents can represent people in civic systems.
The authors say the study explores different ways of adding to preference data and develops tests for individual and group preferences.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. This is the authors’ stated aim, not a conclusion about how well the methods work. Addresses: The paper’s stated contribution
Inspect the contribution and sourcesObjection · Provider control creates risks
Ownership and funding risks qualify proposals for personal agents and user control.
Different model ownership and funding arrangements may create different routes to manipulation or control by powerful interests; the authors say none is immune.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. The authors cite proprietary, open-source (software whose code is publicly available), venture-funded, and government-supported models as examples of different arrangements. Addresses: Ownership and support of language models
Inspect the contribution and sourcesImportant limit · Models and code access
Public availability of fine-tuned models bears on model access and control, but forthcoming reproduction code leaves transparency incomplete and does not prove accountability or reproducibility.
The researchers say their models, further trained for specific tasks, are publicly available, while the code for reproducing the results was to be posted later.
Source coverage: Full readable text of the retrieved 37-page research draft shared before formal publication, including references and appendices. Figures are not independently interpreted. The passage distinguishes models said to be available from code whose repository link was still to be updated. Addresses: Availability of study materials
Inspect the contribution and sourcesImportant limit · The study’s scope
Prediction within this group does not establish representation across a broader electorate.
The study asks whether models can learn the preferences of a specific group of 267 participants. It does not aim to estimate what people across Brazil prefer.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. The authors define the 267 participants as the group whose preferences they aim to model. They state that the data cannot support estimates for people across Brazil.
Inspect the contribution and sourcesImportant limit · A limited research aim
This scope distinction qualifies what the prediction work can establish: studying participation tools does not demonstrate accountable civic representation or the wider democratic effects of communication technologies.
The authors limit their aim to exploring language models as tools for participation, rather than studying how communication technologies have changed modern democracy overall.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. 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
Inspect the contribution and sourcesObjection · Some groups may be missing
Challenges the assumption that personalized agents would adequately represent all citizens, especially views underrepresented in training data.
Biases and other limits may make language models worse at representing some political views, especially views of groups missing from training data.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. 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
Inspect the contribution and sourcesImportant limit · Earlier studies used models to imitate judgments and predict election choices
Prior persona-based prediction is relevant background for agent proposals but does not demonstrate accountable representation in real decisions.
Earlier studies cited by the authors reported similar moral judgments from language models and used detailed descriptions of people to predict election choices.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. These findings are attributed to earlier studies, not to the authors’ own experiment. Addresses: Background research on simulated participants
Inspect the contribution and sourcesImportant limit · Not ready for use in real decisions
Directly qualifies proposals for personalized civic agents: prediction experiments do not establish readiness for real decisions, and the assessment is about current models rather than all future systems.
The authors conclude that current language models are not ready for full augmented democracy systems, while describing the area as a worthwhile subject for research.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. 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
Inspect the contribution and sourcesStarting point · The role of intermediaries in democracy
It motivates exploring participation tools that could change the intermediary relationship, without establishing that agents provide accountable representation.
The authors note that many democracies rely on intermediaries between people and political power. They use this as context for questioning that arrangement and exploring other forms of participation.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. This is the authors’ framing of the motivation for exploring participation tools, not a measured finding about how many democracies rely on intermediaries.
Inspect the contribution and sourcesImportant limit · Portuguese prompt example
The example makes concrete how limited participant details and a recorded policy choice were used to represent an individual in a prompt. It reports no check result and does not establish accountable representation.
For a robustness check, the researchers made Brazilian Portuguese prompts describing participant characteristics. The features included age (young or older/“third age”), political ideology (conservative, centrist, or left-wing), urban or rural zone, education (with or without higher education), sex (male or female), and city and state. One example described a 26-year-old liberal man with a master’s degree in Rio de Janeiro choosing a proposal to curb tax evasion over one to expand privatization of state-owned companies and infrastructure concessions.
Source coverage: Full readable text of the retrieved 37-page research draft shared before formal publication, including references and appendices. Figures are not independently interpreted. 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
Inspect the contribution and sourcesSupporting point · Express preferences on individual policies
It directly describes the proposed use of agents to help citizens express policy-level preferences, while its untested status leaves accountability unresolved.
Software programs that act for people could help citizens express preferences on individual policies instead of choosing only among groups of policies offered by parties or candidates.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. 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
Inspect the contribution and sourcesImportant limit · Reasons to explore language models
These are motivations for exploring agents, but they do not establish accountability, control, or reliable representation.
The authors give ease of use, the ability to interact using everyday language, and a market with many providers as reasons to explore language models for these systems.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. These are reasons for exploring the technology; they do not show that models are safe or effective representatives. Addresses: The proposed choice of technology
Inspect the contribution and sourcesImportant limit · Changing option order could change answers
Order sensitivity qualifies the claim that a personal agent reliably reflects its user’s preferences; it does not establish universal failure.
The authors report that model predictions could be unclear and could change when the order of two options was reversed. In some cases, consistency was below 70%.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. 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
Inspect the contribution and sourcesImportant limit · Models can imitate some responses
Prior simulation and vote-prediction results do not show that a personal agent can represent a person accountably in real civic decisions.
Earlier studies cited by the authors found that language models could simulate some opinion-poll responses and predict votes in elections with two choices.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. These findings are attributed to earlier studies, not to the paper’s own experiment. Addresses: Prior research on simulated opinions
Inspect the contribution and sourcesStarting point · Bundles simplify choice but may not fit all
The mismatch risk supplies context for proposals to support individual-policy choices, while leaving their accountability and representation unresolved.
The authors say parties and politicians have incentives to adapt policy packages to constituents, but there is no guarantee those packages best satisfy every citizen’s preferences.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. 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.
Inspect the contribution and sourcesObjection · Reasons for choices are unclear
Limited explanation challenges confidence that personal agents can be meaningfully assessed or held accountable, without showing that every use is impossible.
The authors say there is not yet a good way to explain why a language model chooses one policy proposal over another.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. 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
Inspect the contribution and sourcesSupporting point · A proposal for AI helpers tailored to individuals
Provides a proposed empirical basis for personalized helpers, but expressly does not demonstrate safe or effective civic representation.
The authors interpret the LoRA result as support for making AI helpers tailored to individuals by adjusting a model through fine-tuning. However, the reported comparison only tests how well predictions match the study participants' recorded preferences.
Source coverage: Full readable text of the 37-page research draft, including references and appendices. Figures are not independently interpreted. A better match on this study's measure supports the authors' proposal for AI helpers tailored to individuals, but does not by itself show that such helpers work safely or well in real civic decisions. Addresses: The authors' interpretation and its limits as evidence
Inspect the contribution and sourcesImportant limit · More limitations may remain
The proposal’s listed concerns should not be treated as exhaustive; further challenges may remain.
The authors say the limitations they list are only some of the many challenges their approach may face.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. The authors explicitly say their list of limitations is not complete. Addresses: The authors’ qualification about the approach’s limitations
Inspect the contribution and sourcesObjection · Accuracy and bias varied
Uneven accuracy and a reported model tendency can complicate personal representation, but do not establish effects for every model or population.
The authors report that their study found better prediction accuracy for some participant groups and that some models tended to favor women.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. 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
Inspect the contribution and sourcesObjection · Accuracy varied across groups
Uneven group accuracy qualifies the prospect that individual agents can represent all users equally well; it does not establish actual overrepresentation.
The authors found that their models predicted pairwise political preferences more accurately for participants who described themselves as younger, liberal, and more educated.
Source coverage: Full readable text of the 37-page research draft, including references and appendices. Figures are not independently interpreted. 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
Inspect the contribution and sourcesImportant limit · Did changing the setting affect model outputs?
Sensitivity tests for two models are relevant context for the reliability of predictions used by personal agents, but the supplied passage gives no results and establishes neither robustness nor sensitivity.
The researchers changed the temperature setting, which affects variation in a model’s predictions, for GPT-3.5 Turbo and Mistral 7B to see whether it affected their outputs. The supplied passage contains no results from those tests.
Source coverage: Full readable text of the retrieved 37-page research draft shared before formal publication, including references and appendices. Figures are not independently interpreted. The appendix identifies two models tested, but only the figure caption is included and no results are described. Addresses: Temperature sensitivity checks
Inspect the contribution and sourcesImportant limit · Age and sex results varied
The disparities qualify proposals for individual agents to represent people, without showing that every model or person is affected alike.
Age comparisons used participants in the younger and older quartiles. Mistral 7B, Falcon 7B, and Gemma 7B showed a slight, statistically significant tendency to predict younger-quartile participants more accurately. Sex-related accuracy varied by model: LLaMA-2 7B, Mistral 7B, and Falcon 7B were more accurate for female participants, while LLaMA-3 8B and Gemma 7B were more accurate for male participants.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. The age result applies to comparisons between the youngest and oldest quarters of participants, not to every possible age group. The sex results also differed across models.
Inspect the contribution and sourcesImportant limit · Possible ways to draw out preferences
This describes a possible mechanism for preference estimates, but does not show that an agent faithfully represents an individual or can be held accountable.
The authors suggest that prompts or additional training might draw out preference patterns represented in language models’ training text.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. 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
Inspect the contribution and sourcesImportant limit · Individual prediction results
Prediction accuracy on held-out recorded choices does not demonstrate that personalized agents can represent people accountably in real decisions.
In a test trained on 50% of participants and evaluated on choices between two options from the other 50%, LLaMA-2 had 76.68% accuracy (reported uncertainty: ±0.0014), GPT 3.5 Turbo 70.4% (±0.0013), Mistral 7B 69.4% (±0.0017), Falcon 7B 69.3% (±0.0014), and BERT SQuAD 50.83% (±0.0015). The authors describe BERT SQuAD’s result as no better than random choice.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. The authors report these accuracy figures for a test set made up of half the participants, after training on the other half. They estimate 95% confidence intervals, ranges used to show uncertainty around a result, with 100 bootstrap iterations. Bootstrapping means repeatedly resampling the data.
Inspect the contribution and sourcesImportant limit · Accuracy differences by education and political views
Group differences qualify claims of uniform representation by personalized agents; the result does not itself show a particular group is overrepresented.
For all six models, the tests found evidence of differences in average accuracy between participants with and without a college education. For all six, the tests also found evidence of differences between liberal and conservative participants, and between liberal and centrist participants. The table marks these results with ***, its symbol for results that meet the 1% significance threshold. This means the results would be unusual if the group averages were actually equal, under the test’s assumptions.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. Passage 44 reports results that meet the table’s significance threshold 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 political-view comparisons
Inspect the contribution and sourcesImportant limit · Differences by political view and education
Uneven accuracy qualifies the proposal that individual agents represent their users, without establishing a population-wide pattern.
Across the six models, the authors report higher accuracy for liberal participants than for conservative and centrist participants, and higher accuracy for college-educated than for non-college-educated participants.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. 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.
Inspect the contribution and sourcesImportant limit · Accuracy gaps raise a concern about whose views models reflect
The concern bears on whether personal agents represent different citizens adequately, but remains a concern rather than a demonstrated representation outcome.
The authors connect differences in accuracy across participant groups to a wider question: might language models reflect some parts of the public more than others?
Source coverage: The full readable text of the 37-page research draft, including references and appendices. Figures are not independently interpreted. The passage links the observed accuracy differences to a concern about whose views models may reflect well. It does not show that the models actually give any group’s views greater weight.
Inspect the contribution and sourcesImportant limit · Preferences beyond political labels
Capturing details beyond a simple political label is relevant to personalized representation, but does not establish full or accountable representation.
The authors interpret the individual prediction result as showing that models capture preference details beyond what the political-alignment rule predicts.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. 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.
Inspect the contribution and sourcesImportant limit · Code and model availability
Model availability may enable some scrutiny, while planned code release leaves reproducibility unresolved; neither establishes accountability or effective control.
The appendix says the code for reproducing the results was to be posted on GitHub, while the models further trained for specific tasks were publicly available there.
Source coverage: Full readable text of the retrieved 37-page research draft shared before formal publication, including references and appendices. Figures are not independently interpreted. 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
Inspect the contribution and sourcesImportant limit · Switching providers may help
Provider choice is a proposed governance safeguard for agents, not proof that capture or manipulation is prevented.
The authors argue that competition among providers may reduce manipulation or control by powerful interests because people can switch providers.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. The paper presents provider choice as a possible protection, while also warning that competition does not remove risks of control by powerful interests or manipulation. Addresses: Provider competition as a possible safeguard
Inspect the contribution and sourcesImportant limit · Some reasoning tasks are difficult
Preference-prediction ability would not establish competence on every reasoning task relevant to civic decisions.
Language models may perform poorly on some reasoning tasks, including mathematics and reasoning about space. This could matter when a decision depends on those skills.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. This is a task-specific caution, not a claim that models fail at every policy task. Addresses: Limits on using models in decision systems
Inspect the contribution and sourcesImportant limit · Models usually beat a simple prediction rule in this study
Better prediction than this rule does not itself establish accountable representation in actual civic decisions.
The models were more accurate than the party-based rule in 21 of 24 comparisons. The exceptions were cases where liberals chose proposals from Bolsonaro’s program for Mistral, Falcon, and Gemma. For participants who identified as left-wing and chose proposals from Lula’s program, LLaMA-2 7B was 15 percentage points more accurate than the rule; all six models reached at least 80% accuracy, compared with 71% for the rule. For choices between proposals from the same candidate’s program, statistical tests found evidence that the models could predict choices, with 65% to 77% accuracy. Overall, the authors conclude that models adapted using LoRA, a method for training a model for a task, predicted individual preferences better than predictions based only on participants’ stated political orientation.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. The results favor the models in most reported comparisons, but include specific exceptions. The authors also report significant predictive power for choices between proposals from the same candidate’s program. Their overall conclusion is limited to the study’s models, participants, proposals, and comparison methods.
Inspect the contribution and sourcesImportant limit · Model choices are hard to explain
Limited explainability qualifies whether users can understand and assess a personalized agent’s choices; this statement does not propose a remedy.
The authors say there is no adequate way to explain why a language model selects one proposal over another.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. The passage identifies the lack of understanding of model choices as a limitation of the proposed approach. Addresses: The authors’ stated limitation
Inspect the contribution and sourcesStarting point · An earlier idea about cooperation
The cited cooperation idea provides historical context for personalized agents as participation partners, but does not establish accountable representation in real civic decisions.
The paper recalls Joseph Licklider’s idea that people and computers working together might make decisions better than either could alone.
Source coverage: Full readable text of the retrieved 37-page research draft, including references and appendices. Figures are not independently interpreted. The authors cite this as background for their interest in people and computers working together; their study does not test Licklider’s claim.
Inspect the contribution and sourcesImportant limit · Training data and transparency matter
Model development and data governance qualify claims that personal agents would be under citizens’ control; transparency is proposed, not established as a sufficient safeguard.
The authors warn that some models’ training data is privately controlled, with access that may be limited, and could be open to manipulation. They call for serious thought about rules for open code and open data to make development more transparent.
Source coverage: Full readable text of the retrieved 37-page research draft, a research draft shared before formal publication or peer review, including references and appendices. Figures are not independently interpreted. 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
Inspect the contribution and sourcesAnswered · How possible does the author think each form of partnership with algorithms is?
The feasibility comparison is relevant context for the narrative’s treatment of personal-agent proposals and their constraints; the question does not establish that either proposal is workable.
How possible does the author think each form of partnership with algorithms is?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Would people control the AI helpers they use?
Dependence on an agent’s makers may concentrate authority rather than establish personal control.
Would people control the AI helpers they use?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Does the author reject partnership with technology altogether?
The answer does not establish that partnerships are accountable or ready for civic use.
Does the author reject partnership with technology altogether?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Can individuals gain the same benefits from data analysis as organizations able to generate and control large datasets?
Unequal control of large datasets limits the proposal’s data advantage, though the evidence does not test a civic agent.
Can individuals gain the same benefits from data analysis as organizations able to generate and control large datasets?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Could an AI model help represent an individual or a community?
Asks whether AI could represent an individual or community, a proposal distinct from validated accountable representation.
Could an AI model help represent an individual or a community?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesSupporting point · A wide range of policies
The example concerns representing fragmented political views through an AI-associated party, relevant to the narrative’s concern with representation beyond preference matching and the accountability of such arrangements.
The Synthetic Party aimed to reflect voters who felt unrepresented through a wide range of policies, including policies that could contradict one another. The chapter describes this as an effort to represent their varied and divided views.
Source coverage: Complete Augmented Deliberation chapter from the book that is updated over time. Other chapters are not included. The passage says the party’s policies covered a wide range and sometimes contradicted one another, in an effort to reflect views among voters who felt unrepresented. Addresses: The Synthetic Party’s proposed use of AI for political representation
Inspect the contribution and sourcesSupporting point · AI models as possible ways to represent people
The proposal directly concerns individual or collective AI representation, while leaving questions of control, accountability, and reliable representation unresolved.
AI text models trained on a person’s writing might imitate that person’s ideas and style. Models trained on a community’s writing might also serve as a way to represent the community.
Source coverage: Complete Augmented Deliberation chapter from the book that is updated over time. Other chapters are not included. The chapter presents models as possible aids, additions, or checks on a human representative’s choices. Addresses: Representation only by a human speaking for a group
Inspect the contribution and sourcesSupporting point · A proposed way to represent people who are disengaged from politics
The proposal concerns AI-mediated representation of disengaged voters, making questions of representation and authority relevant; the stated seat share is only an aim.
The Synthetic Party trained its model on material from small parties. It aimed to make AI politics genuinely democratic rather than focused on appealing to ordinary people against established groups. It proposed using voter disengagement to set the probabilities for allocating representative seats. Its hoped-for share was about 20% of voters, or roughly 20–36 seats in Denmark’s 179-seat parliament.
Source coverage: Complete Augmented Deliberation chapter from the book that is updated over time. Other chapters are not included. The chapter describes the proposed seat allocation as a hypothetical aim, not an achieved result. It links the proposed seats to voter disengagement, meaning less interest or involvement in politics, and to material from small parties. Addresses: The Synthetic Party’s use of AI in politics
Inspect the contribution and sources