With colleagues at CPL, I designed an algorithm that uses LLMs to automatically mask race-related information in police reports. Prosecutors then review these redacted reports, helping them avoid implicit bias when deciding whether to charge or dismiss a case. After we ran pilots at the San Francisco and Yolo District Attorney’s offices, California passed a law requiring prosecutors across the state to adopt our intervention. We are now studying the impacts of blind charging with a randomized controlled trial. Blind charging has been covered in numerous press articles. Learn more at blindcharging.org.

Many studies have framed algorithmic fairness as a mathematical problem, proposing axiomatic constraints without fully considering the consequences of these constraints. My coauthors and I devised a new approach that uses contextual bandits and convex optimization to achieve equitable outcomes. We demonstrate the advantages of this approach using data from the Santa Clara Public Defender in a paper in Management Science.

Failing to appear in court can land a person in jail and cause them to lose their job or housing. But many people fail to appear (FTA) simply because they forget about their court date. In a randomized controlled trial at the Santa Clara Public Defender’s Office, we found that text message reminders reduced FTA-related jail involvement by over 20%. Our findings extend previous studies which found that text message reminders can help people show up to court. We’re now testing whether the standard consequences-focused reminder hurts or helps certain clients, and whether monetary assistance can help clients overcome financial barriers to court attendance.

The last few years have seen an explosion in research on how to constrain algorithms to avoid biased decision-making. My colleagues and I wrote a short guide for Nature Compuational Science that synthesizes this research, illustrates drawbacks to several widely cited approaches, and outlines practical steps people can take in their quest for equitable algorithms.

In a paper, my colleagues and I described how data analysis can identify and measure racially disparate impacts in police stop policies, complementing other research which investigates whether bias exists in individual stops. We gave applied examples from a handful of major cities across the U.S., including Nashville, New York, Chicago, and Philadelphia. Our paper was published in the University of Chicago Law Review, and I wrote a Twitter thread about it here.

In a Washington Post piece, I argued (with colleagues from CPL) that the dramatic—but temporary, and patchwork—criminal justice reforms enacted in response to COVID-19 should be made permanent and expanded across the country.

I wrote a briefer on the potential advantages and risks of using risk assessment instruments in criminal justice settings for the Brookings Institute’s “AI and Bias” series on fairness in algorithms.

I helped design and deploy a tool called Patternizr that helps speed up the investigation of serious historical crimes. My colleague and I described our approach in a paper in the INFORMS Journal of Applied Analytics, including a first-of-its-kind analysis from a police department that demonstrates the tool does not disproportionately recommend any specific race group. Patternizr was featured in several articles.

I worked with colleagues at CPL on a project with the city of Nashville to demonstrate that traffic stops were an ineffective tool for fighting crime. Since the release of our report, the city’s police department has reduced the use of traffic stops by 70%—an almost 90% reduction from their peak.

I designed an algorithm for the NYPD that looked for crimes which were misclassified as felonies or misdemeanors. Likely misclassifications were sent to an internal team for auditing and correction. I presented my approach at NYU’s Tyranny of the Algorithm? Predictive Analytics & Human Rights conference.