Using AI to Redact Race-Related Information from Police Reports

• In Event: At the Frontier: Exploring Applications of Large Language Models in Crime Research and Public Policy

Abstract

In 2022, California’s state legislature passed AB-2778, requiring prosecutors across the state to implement a new “race-blind” charging process by the beginning of 2025. In race-blind charging, prosecutors begin their review of a new case by reading documents where all race-related information has been redacted (where race-related information includes not only explicit mentions of race, but also close proxies like physical descriptions or names) and recording a race-blind decision about whether to charge the case. We designed, developed, and deployed an algorithm that uses generative artificial intelligence and large language models to conduct these redactions automatically. This algorithm is now used by the majority of prosecutors in California to comply with the requirements of AB-2778. In this study, we describe the function, design, and performance of the algorithm. This includes key tradeoffs that ensure the algorithm is both secure in sensitive environments and accessible to local information technology staff. We show that despite these limitations, our algorithm reliably redacts race-related information and makes it difficult to guess an arrestee's race. We also discuss preliminary findings from our evaluation of this new mandate's impact on charging decisions. We conclude with some ideas for future extension of this work.

Authors

• Alex Chohlas-Wood, New York University

• Joe Nudell, Harvard University

• Julian Nyarko, Stanford University

• Sharad Goel, Harvard University