Legal reformers, for repeal campaigns and litigation, identify discriminatory laws through painstaking manual searches. This approach has proven effective at the federal and state levels, but becomes infeasible for local law owing to its scale. We make three contributions. First, we assemble the largest corpus of U.S. local codes and charters, spanning 9,623 jurisdictions and covering approximately 75% of the U.S. population. Second, we introduce a large language model (LLM)-assisted pipeline that combines (i) high-recall candidate identification with (ii) soft legality ratings, prioritizing provisions for human review. We detect laws that explicitly reference protected categories and impose group-specific treatment, and apply few-shot, chain-of-thought prompting to assign priority ratings for expert review. We find that this framework (1) achieves high recall (0.91) and precision (0.91) on a held-out, curated validation set of historical and expert-identified laws, but (2) is less reliable in drawing substantive assessments of antidiscrimination law compared to human reviewers. Third, we show that this system uncovers a wide range of discriminatory provisions, including poll taxes, race and gender-based voting rules, gendered occupational restrictions, and citizenship-based licensing exclusions. Together, these results demonstrate the potential of LLM-assisted workflows to scale review and reform of local law.
Read the associated policy brief here.