The paper asks what happens when an LLM generates a person's ballot from demographic, political, and preference traits, and whether voting design can absorb the resulting errors. Its strongest result is not that AI represents individuals accurately: it performs very poorly on complex ballots. Mean individual human-AI consistency is 5.68% for the actual Aarau vote and 28.005% for its hypothetical survey, versus 84.5% in binary US elections. The useful contribution is different: after thousands of ballots are aggregated, some winner sets are much more stable than the individual predictions, and that stability depends strongly on the aggregation rule.
The study combines three observed settings. ANES supplies the 2012, 2016, and 2020 US presidential elections: the paper begins with 20,650 respondents, retains 17,010 complete profiles, and uses single-choice ballots with majority aggregation. City Idea, a 2023 participatory-budgeting campaign in Aarau, supplies a 3,314-person survey over five projects using single choice, approval, and score ballots, plus an actual 1,703-citizen vote over 33 projects using cumulative ballots; 505 voters have linked traits for emulation. The design adds correctly to 21, 207, and 135 emulated elections, 363 in total.
The study labels six LLMs, GPT-4o Mini, GPT-3.5, GPT-3, DeepSeek R1, Gemini 1.5 Flash, and Llama 3 8B, and one predictive neural network. Prompts are zero-shot, provide election context and personal traits, and request step-by-step reasoning. The supplement says temperatures from 0.4 to 0 were tested with 20 runs per setting and mean consistency across runs. It does not provide a manifest linking each result to a checkpoint, provider, endpoint, parameters, temperature, repetition, exact prompt, parser, or failure. The June-November 2023 and April-August 2025 collection windows are not mapped model by model. Generic model labels therefore do not identify reproducible systems.
The main metric counts relative matches in Condorcet pairwise comparisons between two ballots. Cross-format comparisons convert scores and cumulative votes to binary representations and exclude ties; collective outcomes become winner/loser vectors. This is a defensible ordinal-similarity measure, but it does not validate shared reasons, values, or a voter's mandate. Conversion discards intensity and ties and can change the denominator. In particular, the method of equal shares often funds more low-cost projects than utilitarian greedy, mechanically increasing overlap between winner sets. The authors recognize this: equal shares and Phragmén average 69.9% collective consistency, 31.6 percentage points above greedy, but the advantage falls to 22.8 points after equalizing winner counts.
Among the LLMs, GPT-4o Mini has the highest individual consistency and is reported 4.85 points above GPT-3.5 and 7.85 above Llama 3 8B. The neural-network baseline is 1.7 points above the LLM average individually and 2.9 collectively. Human cross-format consistency is 97.1%; model values fall to 74.3% for GPT-4o Mini, 72.1% for GPT-3.5, 76.2% for Llama 3 8B, 71.23% for Gemini, and 68.7% for DeepSeek. Equal shares raises collective cross-format consistency by 12.2 points over greedy. These are descriptions of this simulation, not stable capability rankings of model families.
The low-turnout experiment is retrospective. In a subset of 252 Aarau participants who completed both pre- and post-vote surveys, overlapping groups with low engagement, digital literacy, or trust are treated as likely abstainers. The analysis then removes ballots from people who actually participated and tries to recover the full-turnout result using their profiles and synthetic ballots. With up to 75% of the low-engagement group replaced by GPT-3.5, equal shares recovers up to 50% more consistency than random removal; proportional methods average a 7.53-point advantage, or 6.72 after controlling winner counts. Yet the supplement also shows that equal shares retains roughly 83-84% of winners with 80% simulated abstention. A substantial part of the resilience belongs to the rule itself, not AI representation.
This is not a study of actual abstainers. It does not observe an abstention decision, frustrated intent to vote, or consent to delegate. Targets are selected correlates among observed participants, and the reference outcome uses those same people's real ballots. The evidence therefore does not show that an LLM knows a nonvoter's preference, can recover a real election, or has democratic authorization. The 2.1-point GPT-4o Mini versus GPT-3.5 contrast is also nonsignificant, p=0.092.
Statistical inference requires substantial caution. The paper uses Fisher's method to combine p-values from runs with different hyperparameters. These tests reuse the same voters, alternatives, and outcomes and are dependent; the number and identity of component tests, base statistics, complete grid, exclusion rules, and global correction across many comparisons are not reported. Most claims provide combined p-value thresholds without paired effect intervals. This is descriptive evidence, not independent replication across models or elections.
Causal language is also excessive. A recurrent network predicts discretized consistency from personal traits, while SHAP, LIME, and ablation rank predictive feature contributions. This does not identify causes: there is no intervention, causal graph, identification strategy, or sufficient control of correlated traits. Post hoc mappings to affect, unconscious, time-discounting, or surrogation bias do not use validated instruments for those constructs. The visible materials do not document a reproducible person-level split, leakage controls, external validation, or complete uncertainty for the RNN and fairness pipeline.
Public reproducibility is partial and belongs to an earlier stage. The official Figshare collection, created in June 2024, contains three ANES workbooks with 5,914, 5,272, and 6,031 rows, 17,217 total. They include profiles, prompts, human choice, GPT-3.5 at temperature 0.2, Llama 2 output, and an ML prediction. They do not match the final 17,010 subset and omit Llama 3, GPT-4o Mini, the other models, Aarau, repetitions, scripts, parsing, aggregation, abstention, RNN, figures, and environments. Internal arithmetic raises another flag: 3,640 incomplete profiles times the three named models would be 10,920 representatives, but the paper reports 7,280, exactly two models. No associated public code repository was found.
The public workbooks embed prompt profiles combining race, gender, age, ideology, party, religion, political interest, state, and inferred vote. No obvious direct identifiers were found, but the combination is sensitive and needs explicit minimization, governance, and linkage-risk documentation. The authors do discuss autonomy, privacy, consent, and accountability and do not simply advocate replacing citizens.
A faithful reading preserves an important conclusion narrower than the title. In these retrospective data, LLMs reproduce complex individual ballots poorly, while proportional rules such as equal shares and Phragmén can make aggregate outcomes more stable under synthetic ballots and simulated turnout loss. Aggregation design is therefore a relevant safeguard to investigate. The study does not establish faithful representation of abstainers, demographic fairness of AI agents, causal democratic resilience, causal cognitive biases, or full reproducibility of the final experiment.