AI Alignment for Digital Democracy: Fairer Voters Representation against LLM Biases

Yi Qi; Srijoni Majumdar; Evangelos Pournaras

cs.CYOct 23, 2026

AI Alignment for Digital Democracy: Fairer Voters Representation against LLM Biases

Authors: Yi Qi, Srijoni Majumdar, Evangelos Pournaras

Organizations: University of Leeds Leeds, United Kingdom

Abstract

AI representation is emerging in digital democracy, assisting in voter emulation and decision-making. Large language models (LLMs) behave differently depending on the model, the prompt context, and the voter information they receive, representing some groups more accurately than others and potentially making the emulation unfair to certain groups. We investigate demographic biases in LLM-emulated electorates: how accurately LLMs represent individual votes, whether some voter groups are represented more accurately than others, and whether AI alignment can reduce these disparities without sacrificing overall accuracy. Using three U.S. presidential elections, we compare a closed-source and an open-source LLM across voters' demographic features and test AI alignment strategies toward fairer voter representation. Removing gender from the voter profile reduces the largest group error of the models by 22-42%, and removing education reduces it by 64%, whereas removing self-reported race increases the mean group error by over 40%, all without lowering accuracy. Common alignment strategies, such as repeated queries or fairness instructions, do not reduce these errors and can double them.

Figures & tables

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.

	Per-group comparison	Fairness metric	Race/ ethnicity	Multiple elections	Open & closed LLMs	Voter emulation
Argyle ( Argyle et al., 2023 )	$\triangle$	\ding 55	\ding 51	\ding 55	\ding 55	\ding 51
Bisbee ( Bisbee et al., 2024 )	$\triangle$	\ding 55	\ding 51	\ding 51	\ding 55	\ding 51
von der Heyde ( Von der Heyde et al., 2024 )	\ding 51	\ding 55	$\triangle$	\ding 55	\ding 55	\ding 51
Yu ( Yu et al., 2024 )	$\triangle$	$\triangle$	\ding 51	\ding 51	\ding 51	\ding 51
Majumdar ( Majumdar et al., 2026 )	$\triangle$	$\triangle$	$\triangle$	\ding 51	\ding 51	\ding 51
Hall ( Hall et al., 2026 )	\ding 51	\ding 51	\ding 51	–	$\triangle$	\ding 55

Table 1. Comparison of closely related work across six dimensions of fairness evaluation in AI voter representation.

Dataset	Election	Vote task	$N$	Role
ANES 2012	Obama vs. Romney	Presidential vote	3,793	Temporal validation
ANES 2016	Clinton vs. Trump	Presidential vote	1,978	Temporal validation
ANES 2020	Biden vs. Trump	Presidential vote	5,141	Main evaluation

Table 2. Datasets used in the evaluation

Figure 2. Worst-group error over voters who self-identify with a racial or ethnic group (top) and individual accuracy (bottom) at each level of control, with one run per voter (solid) or five runs at the base level (hollow). Accuracy is stable across levels, while race-targeted control does not reduce the worst-group error: for Claude, prompt-based control raises it in every election.

Figure 3. Signed misrepresentation error (share of emulated Democratic choices minus reported share of Democratic votes) for the voter choices emulated by Claude and by Llama, by self-identified racial group, ANES 2020 (base level, K = 5; bars: 95% intervals). For both models, the share of emulated Democratic choices is furthest from the reported share for voters who self-identify as Native American (about 14 points higher) and within 1 point for voters who self-identify as Black.

Figure 4. Signed misrepresentation error (Share of emulated Democratic choices minus reported share) by group at the base level (

K=5

), with 95% intervals; filled markers exclude zero. For voters who self-identify as male and voters with at most a high school education, the share of Democratic choices emulated by both models is below the reported share in every election, while the errors for racial groups vary across elections.

Figure 5. Gap amplification on ANES 2020 (1 = real group differences reproduced; log scale; 95% intervals). Given only the attribute, both models exaggerate group differences several-fold; the full profile still exaggerates gender and education gaps.

		Worst-group error		Mean group error
1=0
1=0
Model	Left out	Full $\to$ without	Change	Full $\to$ without	Change	Acc. change
Claude	Race	.124 $\to$ .185	$+49\%$	.043 $\to$ .061	$+42\%$	$-0.5$
	Gender	.107 $\to$ .062	$-42\%$	.071 $\to$ .054	$-25\%$	$-0.3$
	Education	.109 $\to$ .039	$-64\%$	.048 $\to$ .021	$-57\%$	$+0.2$

Table 3. Group error when one attribute is left out of the profile (ANES 2020). Negative change = fairer.

					$\varepsilon_{\max}$
1=0
Model	Year	Level	Acc.	Pop.	Race	Gender	Educ.
Claude	2012	Base	.822	.026	.051	.085	.048
		Prompt	.825	.043 ∗	.140 ∗	.083	.091 ∗
		Decomposed	.815 ∗	.050 ∗	.084	.095 ∗	.078 ∗
	2016	Base	.832	.001	.071	.094	.076

Table 4. Accuracy, population-level error, and worst-group error

\varepsilon_{\max}

by attribute at each level of control (

K=1

, same voters). ∗ : differs from base (

p<0.05

, paired bootstrap).

Figure 6. Signed misrepresentation error by group on ANES 2020 at each level of control (

K=1

, same voters). Prompt-based control doubles Claude’s error for voters who self-identify as Native American; activation-level control shifts every Llama group toward the Democratic candidate.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Meaning
$\mathcal{G}$ , $r$	Set of groups under one attribute; a group
$J_{r}$ , $N_{r}$	Voters in group $r$ ; their number
$y_{j}$	Reported vote of voter $j$ ( $1$ = Democrat)
$\hat{y}_{j,k}$	Emulated vote of voter $j$ in run $k=1,\dots,K$
$\hat{p}_{j}$	Share of runs voting Democrat, $\frac{1}{K}\sum_{k}\hat{y}_{j,k}$
$\hat{y}_{j}$	Majority vote, $\mathbb{1}[\hat{p}_{j}>1/2]$

Appendix

Table 5. Notation.

Attribute	Groups
Race	White, Black, Hispanic, Asian/NHPI, Native American, Other
Gender	Man, Woman
Age	18–34, 35–49, 50–64, 65+
Education	High school or less, Some college, Bachelor’s, Postgraduate

Appendix

Table 6. Demographic groups used in the analysis.

$\alpha$	1	2	4	8
Democratic share	0.65	0.73	0.99	1.00

Appendix

Table 7. Overall Democratic share under activation-level control on 200 ANES 2020 voters (uncontrolled Llama: 0.58; reported: 0.58).

Section	Elections	Level of control	Runs
Sec.5.2 Robustness	2012, 2016, 2020	Base	$K=5$
Sec.5.3 Source	2020	Base, three profile variants	$K=1$ or $5$
Sec.6.1 Aggregation	2012, 2016, 2020	Base	$K=5$
Sec.6.2 Control	2012, 2016, 2020	Base, prompt-based, decomposed	$K=1$
Sec.6.4 Prompt vs. activation	2020	Decomposed, activation (Llama)	$K=1$

Appendix

Table 8. Experiments. All use the voters described in Appendix B.1 .

				$\varepsilon_{\max}$
1=0
Model	Year	$K$	Acc.	Race	Gender	Educ.
Claude	2012	1	.822	.051	.085	.048
		3	.822	.051	.086	.048
		5	.823	.051	.085	.047
	2016	1	.832	.071	.094	.076

Appendix

Table 9. Accuracy and worst-group error

\varepsilon_{\max}

by attribute under majority voting over

K

runs (base level).

Method	Prompt Template
Unsteered LLM	You are simulating a survey respondent in a U.S. election study. Given the following voter profile, predict which candidate this respondent would vote for. Profile: {profile_text}. Question: Based on this profile, would this respondent vote for {Candidate A} or {Candidate B}? Answer with only one candidate name.
Naive Prompt Steering	You are simulating a survey respondent in a U.S. election study. Given the following voter profile, predict which candidate this respondent would vote for. Profile: {profile_text}. When making the prediction, consider both individual-level electoral factors such as ideology and political interest, and avoid systematically under-representing the voting preferences of the respondent’s racial or ethnic community. Question: Based on this profile, would this respondent vote for {Candidate A} or {Candidate B}? Answer with only one candidate name.
Decomposed Prompt Steering	You are simulating a survey respondent in a U.S. election study. Given the following voter profile, predict which candidate this respondent would vote for. Profile: {profile_text}. [Shared electoral reasoning block] Consider task-relevant individual-level signals, including ideology, political interest, age, education, and other non-sensitive profile attributes. [Attribute-aware representation block] Separately consider whether the prediction would systematically under-represent the voting pattern of the respondent’s racial or ethnic community. This block should be treated as an independent consideration and should not override the individual-level profile evidence. [Final prediction] Based on both blocks, would this respondent vote for {Candidate A} or {Candidate B}? Answer with only one candidate name.

Appendix

Table 10. Prompt templates for the three levels of control. Placeholders in braces are replaced with the voter’s profile and the two candidates.

Model	Year	Grouping	$\varepsilon_{\max}$	Worst subgroup	Signed ( $n$ )	Sig.
Claude	2012	Race	.051	White	$-.051$ (2437)	2/5
		Race $\times$ gender	.139	White men	$-.139$ (1205)	3/9
		Race $\times$ age	.118	Other, 35–49	$-.118$ (34)	3/14
		Race $\times$ education	.149	Hispanic, Bachelor’s	$+.149$ (87)	3/14
	2016	Race	.071	Asian/NHPI	$+.071$ (42)	0/5
		Race $\times$ gender	.175	Hispanic men	$-.175$ (80)	4/8

Appendix

Table 11. Worst-group error

\varepsilon_{\max}

for race and its intersections (base level,

K=5

). “Worst subgroup” gives the subgroup with the largest error, its signed error (positive: share of emulated Democratic choices above the reported share), and its size. “Sig.” counts subgroups whose 95% interval excludes zero, out of all subgroups with at least 30 respondents.

Model	Left out	Metric	Full	Without	Difference [95% CI]	Change	$p$
Claude	Race	Worst-group error	.124	.185	$+.061$ [ $-.050$ , $+.153$ ]	$+49\%$	.294
		Mean group error	.043	.061	$\mathbf{+.018}$ [ $+.004$ , $+.043$ ]	$\mathbf{+42\%}$	.008
		Accuracy	.847	.842	$-.005$ [ $-.011$ , $.000$ ]		.086
	Gender	Worst-group error	.107	.062	$\mathbf{-.045}$ [ $-.054$ , $-.036$ ]	$\mathbf{-42\%}$	$\mathbf{<.001}$
		Mean group error	.071	.054	$\mathbf{-.018}$ [ $-.031$ , $-.004$ ]	$\mathbf{-25\%}$	.020
		Accuracy	.847	.844	$-.004$ [ $-.011$ , $+.003$ ]		.328

Appendix

Table 12. Change when one attribute is left out of the voter profile (ANES 2020). Differences are “without” minus “full”; negative error differences mean fairer representation. Bold:

p<0.05

Demographic Bias in Language Models

Demographic bias in language models studies how identity cues shape model outputs, decisions, and representations, and how to detect and reduce resulting disparities.

See all 135 papers

Diagnosing Korean-Language LLM Political Bias via Census-Grounded Agent Simulation
May 18, 2026Sungwoo KangPolitical Bias in Language Models Language Model Calibration
Department of Electrical and Computer Engineering Korea University, Seoul, Republic of Korea
Large Language Models as Implicit Sociological Models: Reconstructing Voting Behaviour from Sociodemographic Profiles
Aug 16, 2026Roman Neruda, Martin Bakoš, Josef Šlerka +3Computational Social Science Large Language Model-Based Social Simulation
Institute of Computer Science, Czech Academy of Sciences, Prague, Czech Republic. · Faculty of Arts, Charles University, Prague, Czech Republic. · distil labs, Berlin, Germany.
Who Would You Vote For? Auditing Political Alignment in LLMs: An Italian Case-Study
Aug 12, 2026Simone MungariLLM Auditing Political Bias in Language Models
Revelis s.r.l.

May 18, 2026J/K move · Enter open · S save

Diagnosing Korean-Language LLM Political Bias via Census-Grounded Agent Simulation

Sungwoo Kang

Department of Electrical and Computer Engineering Korea University, Seoul, Republic of Korea

Political Bias in Language Models Language Model Calibration Language Model Bias Evaluation

Open paper AskPDF arXiv

Large language models (LLMs) exhibit systematic political biases in voter simulations, but their underlying mechanisms and cross-lingual generalizations remain poorly understood. We introduce Dynamo-K, a census-grounded simulation framework evaluating Korean-language LLM political behavior across four models on six Korean elections (2017-2025). Using this framework, we identify three systematic failure modes: (1) progressive bias in moderate agents, where explicit mitigation reduces Mean Absolute Error (MAE) by 5.2 times; (2) model-dependent third-party salience collapse, distinguishing between salience failure and decision bias; and (3) regional polarization collapse, where models bidirectionally under-predict historical party strongholds. To address these failures, we demonstrate that scenario reframing recovers 62% of 2017 MAE by restoring third-party visibility. Furthermore, we introduce a learned reweighting adapter that successfully calibrates opposing-valence models without relying on candidate names at train or test time. Validating our diagnostic framework, Dynamo-K accurately predicts 3/3 presidential winners - including a 2.1%p MAE on the highly contested 0.73%p-margin 2022 race - and correctly identifies the dominant party in a held-out local election. The pipeline is open-source and provides a scalable, cost-effective method for diagnosing LLM political behavior.

Abstract

Demographic Bias in Language Models

Explore similar work

Diagnosing Korean-Language LLM Political Bias via Census-Grounded Agent Simulation

Large Language Models as Implicit Sociological Models: Reconstructing Voting Behaviour from Sociodemographic Profiles

Who Would You Vote For? Auditing Political Alignment in LLMs: An Italian Case-Study