What Can AI Tell Us About NYPD Street Stops?
RESEARCH STUDY
What Can AI Tell Us About NYPD Street Stops?
Using Natural Language Processing and Machine Learning to
Analyze Investigative Encounters and Consent Searches from
Police Body-Worn Camera Footage
Rob Voigt, Nicholas P. Camp, Dan Sutton, Jennifer L. Eberhardt
Published: April 2026
Download Report
Executive Summary
EXECUTIVE SUMMARY
In 2013, the U.S. District Court for the Southern District of New York found that the NYPD’s stop-and-frisk practices violated the Fourth Amendment, which requires stops to be based on reasonable suspicion, and the Fourteenth Amendment through a pattern of racial profiling during stops. The Court appointed an Independent Monitor to oversee reforms to the Department’s policies, training, supervision, and documentation of investigative encounters.
Body-worn cameras, required by the Court’s remedial orders, now provide an objective record of what officers actually say during encounters—information that is largely invisible in stop reports and other administrative records, but essential for assessing constitutional compliance. In 2021, the Court approved this study by a Stanford-affiliated research team, which uses AI tools and techniques—including machine learning and natural language processing—to computationally analyze BWC recordings and identify key indicators of constitutional compliance in what NYPD officers say and how they say it during encounters with civilians.
This study focuses on two areas where AI and language-based methods provide the clearest and most meaningful insights for compliance monitoring: the NYPD’s compliance with New York’s De Bour framework for classifying police interactions and the NYPD’s consent search practices.
First, the study evaluates the NYPD’s compliance with New York’s De Bour framework for classifying police interactions, with a specific emphasis on Level 3 stops and detentions that require reasonable suspicion. Second, the study analyzes the NYPD’s consent search practices, where the Fourth Amendment requires that consent be voluntary and depends in significant part on the language officers use.
Automatic speech recognition converts body-worn camera audio into text transcripts; natural language processing and machine learning models then examine these transcripts to identify patterns in officer language. To assess compliance with the Fourteenth Amendment, the study analyzes differences in officer language during encounters with civilians of different races and ethnicities, from low-level interactions to stops and searches.
The study’s key findings include:
- Machine learning models can reliably distinguish De Bour encounter levels and the study shows these models could be used to select potentially problematic footage for human review and substantially raise identification rates of undocumented stops. Application of these models to a broader spectrum of BWC footage would likely uncover large numbers of unreported stops.
- The prevalence of explicit consent language during consent searches is consistently low: the word “search” appears in only approximately 46% of consent search encounters, “consent” in approximately 13%, and confirmatory questions in approximately 21%. Interactions in which all three key terms appear—consistent with the phrasing set out in the NYPD Patrol Guide—occur in only 3.2% of interactions. These patterns raise concerns under Fourth Amendment requirements, the Right to Know Act, and NYPD Patrol Guide policy.
- Across encounter classification and consent search analyses, the study identifies significant racial disparities in officer language. Relative to White civilians, Level 1 and Level 2 encounters involving Black and Hispanic civilians are linguistically more similar to Level 3 stops. Detailed analysis of officer language identifies contributing factors including the increased use of direct commands, casual language, and profanity with Black and Hispanic civilians.
This report demonstrates that computational analysis of body-worn camera footage can meaningfully strengthen constitutional compliance monitoring at the scale urban policing requires. The NYPD records millions of investigative encounter videos each year, yet current monitoring efforts necessarily capture only a small fraction of officer-civilian interactions. The findings suggest practical applications across several areas of NYPD operations, including officer training, supervisory tools, and policy development.
With access to more recent and comprehensive data, these methods could support the Court, the Monitor, and the Department in evaluating the impact of ongoing reform efforts. The Stanford team has applied this approach in two large cities in California, analyzing over 1.3 million videos representing more than 300,000 hours of interactions—demonstrating the potential for computational analysis to support compliance assessment at department-wide scale.
American police departments record millions of body-worn camera videos each year, but even well-resourced review efforts can examine only a small fraction of them. Important questions about how police interact with their communities remain difficult to answer. Advances in AI, paired with appropriate safeguards and collaboration among researchers, departments, and those responsible for oversight, make it possible to analyze this footage in ways that move beyond limited audits toward more comprehensive and systematic assessment. This is creating opportunities to identify problems earlier, measure whether reforms are working in practice, and build the kind of accountability that strengthens both constitutional compliance and public trust.