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

What it claims

Pournaras, Qi and Majumdar (University of Leeds, arXiv October 2026) show that LLM-emulated U.S. electorates can reach about 85% individual vote accuracy while one demographic group's emulated Democratic share is off by 10 points or more. The two models are named in the contributions: Claude (closed-source) and Llama (open-source). Removing gender from the voter profile cuts the largest gender-group error by 42% for Claude and 22% for Llama. Removing education cuts Claude's largest education-group error by 64% (p < 0.01). Removing self-reported race raises mean group error by over 40% (p < 0.05). None of these removals lowers accuracy. Repeated queries, aggregation, fairness instructions and activation-level control do not reduce misrepresentation. A fairness instruction can double one group's error.

How it works

The data are every eligible respondent in ANES 2012, 2016 and 2020. Each profile includes fields such as ideology, age, education and political interest, plus a group attribute. The main grouping is self-reported race or ethnicity. Votes are binary (Democrat or Republican). Third-party voters and non-voters are excluded.

The model is queried an odd number of times K per voter. The emulated vote is the majority (self-consistency decoding). Per-voter consistency is max(p, 1 − p), where p is the share of runs predicting Democrat. Group error is the absolute gap between a group's emulated and reported Democratic share, in vote-share points. The main fairness metric is the largest gap across groups. Mean group error and population-level error are also reported. A group counts as "misrepresented" when the 95% confidence interval of its signed error excludes zero, and group sizes are reported because small groups give noisy estimates.

Interventions come at three levels: the base model, a prompt-based fairness instruction, and a decomposed prompt that separates vote-relevant reasoning from group representation. Llama also gets activation-level control.

Two arguments explain why repeated sampling fails. Majority voting removes random variation but keeps errors the model makes in every run. A change in a voter's vote probability only affects the group share if it crosses 0.5, and then it moves the share by a full 1/N. A short lemma shows that averaging probabilities instead of majority votes bounds the error of any convex combination of predictors by the worst predictor's error. No such guarantee holds for majority votes.

The authors also argue that standard fairness definitions (equalized odds, similar treatment across groups) do not fit voter emulation. Groups really do vote differently, so equalizing predictions across groups can increase misrepresentation.

Limits

The fetched text stops at the start of the experimental setup. It does not give model versions, the value of K, temperature, prompts, per-group result tables, sample sizes, or details of the activation steering. The "10 points or more" and "near 85%" figures come from the contributions list, not from shown tables. The setting is binary two-party U.S. presidential vote choice only. The race grouping is the primary analysis; other attributes are described as possible extensions. No code or data release is mentioned in the fetched portion.