Figure 1. Voter misrepresentation in AI representation. High individual accuracy (85%) and a small overall error (3 points) can coexist with a large error for one group (Group C, 14 points), reversing its ordering relative to another group.Large language models (LLMs) are increasingly used to emulate human respondents by grounding predictions on demographic, ideological, and contextual profiles ([Argyle et al., 2023](https://arxiv.org/html/#bib.bib2); [Zhang et al., 2024](https://arxiv.org/html/#bib.bib4); [Li et al., 2025a](https://arxiv.org/html/#bib.bib3); [Majumdar et al., 2026](https://arxiv.org/html/#bib.bib1); [Ziems et al., 2024](https://arxiv.org/html/#bib.bib20)). In electoral settings, an LLM can predict a voting choice for each real voter profile and aggregate these predictions into a synthetic electorate. We refer to this process as *AI representation of voters*. It is used to pilot and complement pre-election polling and forecasting ([Cerina and Rouméas, 2025](https://arxiv.org/html/#bib.bib39); [Markaki et al., 2026](https://arxiv.org/html/#bib.bib40)), to estimate the preferences of abstaining or hard-to-reach voters who are missing from surveys ([Majumdar et al., 2026](https://arxiv.org/html/#bib.bib1); [Gholamnia et al., 2026](https://arxiv.org/html/#bib.bib41)), to build synthetic populations for policy and social-science research ([Argyle et al., 2023](https://arxiv.org/html/#bib.bib2); [Zhang et al., 2024](https://arxiv.org/html/#bib.bib4)), and to support deliberation and collective decision-making in digital democracy ([Majumdar et al., 2026](https://arxiv.org/html/#bib.bib1); [Rymon, 2026](https://arxiv.org/html/#bib.bib42)). In each of these uses, AI representation is valuable only if the resulting electorate reproduces aggregate vote shares and individual voters and demographic groups are represented faithfully and as intended. We distinguish *individual fidelity*, where an emulated vote matches the corresponding human respondent’s reported vote, from *aggregate accuracy*, where the simulated electorate reproduces empirical vote-share distributions. These attributes do not align: aggregate agreement can conceal errors that are systematically concentrated within particular voter groups. Prior work on LLM-based opinion simulation shows that model responses to survey and political questions shift with the demographic profile the model is conditioned on, and often align better with some groups’ opinions than others ([Argyle et al., 2023](https://arxiv.org/html/#bib.bib2); [Sun et al., 2024](https://arxiv.org/html/#bib.bib8); [Lee et al., 2024](https://arxiv.org/html/#bib.bib9); [Santurkar et al., 2023](https://arxiv.org/html/#bib.bib18); [Bang et al., 2024](https://arxiv.org/html/#bib.bib34)). Studies that emulate voters specifically report a second problem: the predicted vote totals can miss the real election result, and the predicted differences between groups can be larger than those observed in real voters ([Kreutner et al., 2026](https://arxiv.org/html/#bib.bib5); [Potter et al., 2024](https://arxiv.org/html/#bib.bib36); [Von der Heyde et al., 2024](https://arxiv.org/html/#bib.bib10); [Yu et al., 2024](https://arxiv.org/html/#bib.bib6)). What remains open is whether the error in predicting individual votes is itself unevenly distributed across voter groups, and whether this unevenness distorts the group vote shares of the emulated electorate. 出典: AI Alignment for Digital Democracy: Fairer Voters Representation against LLM Biases(arXiv)
全体の数字の下にある監査
著者らは、2012年、2016年、2020年の米国大統領選挙について、American National Election Studiesの回答者を対象にClaude Sonnet 4.6とLlama-3.1-8B-Instructを検証した。2020年には、両モデルが個々の投票を正しく予測した割合は84.8%と85.8%だった。