A new study published on arXiv details the validation of bc2, an open-source LLM-based algorithm designed to automate race-blind charging decisions in California. The algorithm was used in over 119,000 cases in 2025. Researchers found that bc2 faithfully implements the state's requirements in 96.7% of narratives, significantly improving upon earlier versions and other open-source methods. However, the study also revealed that California's mandate misses key racial proxies like location information, and that bc2's broader redactions eliminate 43.1% of the remaining predictive signal for race. AI
IMPACT This research demonstrates how LLM validation can improve algorithms and help policymakers achieve underlying goals, potentially influencing future AI applications in legal and policy domains.
RANK_REASON The cluster contains an academic paper detailing the validation of an algorithm for a specific policy application. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
