Study: What can AI tell us about NYPD street stops?
Thousands of officer-worn camera recordings found evidence of underreported police stops, troubling racial disparities in officer interactions, and widespread use of unclear language during consent searches, a new study shows.
Researchers at the University of Michigan, University of California-Davis and Stanford University say their findings raise constitutional concerns under both the Fourth and Fourteenth Amendments, involving protection from unreasonable searches/seizures and prohibiting discriminatory practices based on race and ethnicity, respectively.
The report highlights how artificial intelligence could transform police oversight by helping reviewers identify potentially problematic encounters hidden within millions of hours of body-camera footage. The research demonstrates the growing potential for AI-powered analysis to help courts, police departments and municipal governments better evaluate compliance while building greater public trust in law enforcement.
Using machine learning and natural language processing, researchers examined New York Police Department encounters captured on body-worn cameras, looking closely at whether officers followed legal standards governing stops, detentions and consent searches.
Among the study's most significant findings:
- Body-camera recordings could be classified as stops with over 80% accuracy, and underdocumented stops with over 70% accuracy based on language alone.
- Using language models, reviewers could uncover over 50% of undocumented stops identified in manual audits by viewing a fraction (25%) of the footage they normally would.
- Officers frequently relied on indirect or confusing phrases such as "Do you mind if I check?" rather than clearly asking for consent to search.
- The word "consent" appeared in less than 13% of consent-search interactions reviewed.
- Commands and indirect requests appeared more frequently in encounters involving Black and Hispanic civilians.
Nicholas Camp, U-M assistant professor of organizational studies, said these patterns raise questions about whether some civilians clearly understood they could refuse searches and whether certain encounters were documented accurately.
The study stems from reforms ordered after the landmark 2013 federal court ruling in Floyd v. City of New York, in which the U.S. District Court for the Southern District of New York found that the NYPD's stop-and-frisk practices violated constitutional protections against unreasonable searches and racial discrimination.
Following the ruling, the court appointed an independent monitor to oversee reforms involving NYPD training, supervision and investigative encounters. As part of those reforms, NYPD officers began using body-worn cameras, which captured numerous police-community interactions.
"These recordings provide a far clearer picture of officer behavior than written police reports alone," Camp said.
The study, approved by the court in 2021, analyzed more than 1,700 encounters connected to an earlier City University of New York Institute for State and Local Governance review, more than 1,100 additional encounters reviewed by the Monitor team, and nearly 1,800 consent-search encounters from 2023.
AI models developed during the study successfully distinguished lower-level encounters from Level 3 stops-which legally require reasonable suspicion-with accuracy rates ranging from approximately 72% to 91%. Researchers say those tools could help oversight teams identify constitutional concerns faster and more consistently by prioritizing footage most likely to contain problematic interactions.
Researchers emphasized that artificial intelligence is not intended to replace human oversight, but instead serves as a tool to strengthen accountability, improve auditing and support ongoing police reform efforts.
"Our analyses identify troubling patterns in NYPD encounters, but also show a path forward: Body camera footage can be used as data to inform and measure changes in law enforcement," Camp said.
The study's authors also include Rob Voigt, assistant professor of linguistics, UC-Davis; Dan Sutton, director of Justice and Safety, Stanford (Law School) Center for Racial Justice, and Jennifer Eberhardt, professor of organizational behavior and psychology, Stanford University.
