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

arXiv:2610.32780v1 Announce Type: new 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.

LLMs may emulate voters unfairly, representing some groups more accurately than others.

The study identifies demographic biases in LLM-based voter emulation, showing disparities in representation accuracy among groups. While some alignment strategies reduce these biases without harming accuracy, others can worsen fairness. This implies potential harm in digital democracy but also opportunities for fairer AI use, hence a mixed but cautiously optimistic impact on human flourishing.

This is a methodological study exposing real biases in LLM voter emulation with quantified results, a niche but concrete application. The implication is concerning for emerging AI-assisted democratic tools but the scope and current real-world deployment are limited.

LLMs deployed in digital democracy for voter emulation show significant demographic biases, representing some groups more accurately than others. Simple data removal strategies can reduce some biases, but not all. Common alignment strategies are ineffective or counterproductive, indicating a systemic challenge to fair representation in democratic processes.

The paper presents evidence that AI alignment can reduce biases in voter representation. Specific demographic features, when removed, significantly reduce errors. This implies a positive impact on fairness in digital democracy.

SOURCE     arXiv cs.CY (Computers and Society) (tier 1)
   ↓
DOCUMENT   f5550a38-f305-4c88-815d-899c31f82d51
           https://arxiv.org/abs/2610.32780
   ↓
EVIDENCE   3 extracted excerpts
   ↓
MODEL RUN  5 runs, methodology 1.0.1
   ↓
SCORE      -10  (Mixed or uncertain)
   ↓
CONFIDENCE 64%