AI Voter Personas: US 2024

Problem

How can we generate structured, data-grounded responses to new questions without fielding new surveys?

Approach

• Variance-weighted K-means clustering (15 voter segments from 51 policy variables)
• Silhouette and stability diagnostics
• 10-question probabilistic assignment quiz
• Structured LLM conditioning grounded in cluster-level distributions
• Controlled prompting to prevent drift

Validation

• Tested on 200 held-out ANES respondents
• Predictions meaningfully above chance
• Accuracy improved when demographics were included

Limitations

• Smooths over individual inconsistency
• Augments but does not replace real surveys