A Los Angeles Police Department Officer demonstrates the use of a body camera in 2015.

Police Turn to AI to Review Bodycam Footage

Over the last decade, police departments across the U.S. have spent millions of dollars equipping their officers with body-worn cameras that record what happens as they go about their work. Everything from traffic stops to welfare checks to responses to active shooters is now documented on video.

The cameras were pitched by national and local law enforcement authorities as a tool for building public trust between police and their communities in the wake of police killings of civilians like Michael Brown, an 18-year-old Black teenager killed in Ferguson, Missouri in 2014. Video has the potential not only to get to the truth when someone is injured or killed by police, but also to allow systematic reviews of officer behavior to prevent deaths by flagging troublesome officers for supervisors or helping identify real-world examples of effective and destructive behaviors to use for training.

This story was originally published by ProPublica and is republished here under a Creative Commons license.

But a series of ProPublica stories has shown that a decade on, those promises of transparency and accountability have not been realized.

One challenge: The sheer amount of video captured using body-worn cameras means few agencies have the resources to fully examine it. Most of what is recorded is simply stored away, never seen by anyone.

Axon, the nation’s largest provider of police cameras and of cloud storage for the video they capture, has a database of footage that has grown from around six terabytes in 2016 to more than 100 petabytes today. That’s enough to hold more than 5,000 years of high definition video, or 25 million copies of last year’s blockbuster movie “Barbie.”

“In any community, body-worn camera footage is the largest source of data on police-community interactions. Almost nothing is done with it,” said Jonathan Wender, a former police officer who heads Polis Solutions, one of a growing group of companies and researchers offering analytic tools powered by artificial intelligence to help tackle that data problem.

The Paterson, New Jersey, police department has made such an analytic tool a major part of its plan to overhaul its force.

In March 2023, the state’s attorney general took over the department after police shot and killed Najee Seabrooks, a community activist experiencing a mental health crisis who had called 911 for help. The killing sparked protests and calls for a federal investigation of the department.

The attorney general appointed Isa Abbassi, formerly the New York Police Department’s chief of strategic initiatives, to develop a plan for how to win back public trust.

“Changes in Paterson are led through the use of technology,” Abbassi said at a press conference announcing his reform plan in September, “Perhaps one of the most exciting technology announcements today is a real game changer when it comes to police accountability and professionalism.”

The department, Abassi said, had contracted with Truleo, a Chicago-based software company that examines audio from bodycam videos to identify problematic officers and patterns of behavior.

For around $50,000 a year, Truleo’s software allows supervisors to select from a set of specific behaviors to flag, such as when officers interrupt civilians, use profanity, use force, or mute their cameras. The flags are based on data Truleo has collected on which officer behaviors result in violent escalation. Among the conclusions from Truleo’s research: Officers need to explain what they are doing.

“There are certain officers who don’t introduce themselves, they interrupt people, and they don’t give explanations. They just do a lot of command, command, command, command, command,” said Anthony Tassone, Truleo’s co-founder. “That officer’s headed down the wrong path.”

“In any community, body-worn camera footage is the largest source of data on police-community interactions. Almost nothing is done with it.”

For Paterson police, Truleo allows the department to “review 100% of body worn camera footage to identify risky behaviors and increase professionalism,” according to its strategic overhaul plan. The software, the department said in its plan, will detect events like uses of force, pursuits, frisks, and non-compliance incidents, and allow supervisors to screen for both “professional and unprofessional officer language.”

Paterson police officials declined to be interviewed for this story.

Around 30 police departments currently use Truleo, according to the company. In October, the NYPD signed on to a pilot program for Truleo to review the millions of hours of footage it produces annually, according to Tassone.

Amid a crisis in police recruiting, Tassone said some departments are using Truleo because they believe it can help ensure new officers are meeting professional standards. Others, like the department in Aurora, Colorado, are using the software to bolster their case for emerging from external oversight. In March 2023, city attorneys successfully lobbied the City Council to approve a contract with Truleo, saying it would help the police department more quickly comply with a consent decree that calls for better training and recruitment and collection of data on things like use of force and racial disparities in policing.

Truleo is just one of a growing number of such analytics providers.

In August 2023, the Los Angeles Police Department said it would partner with a team of researchers from the University of Southern California and several other universities to develop a new AI-powered tool to examine footage from around 1,000 traffic stops and determine which officer behaviors keep interactions from escalating. In 2021, Microsoft awarded $250,000 to a team from Princeton University and the University of Pennsylvania to develop software that can organize video into timelines that allow easier review by supervisors.

“There are certain officers who don’t introduce themselves, they interrupt people, and they don’t give explanations. They just do a lot of command, command, command, command, command.

Dallas-based Polis Solutions has contracted with police in its hometown, as well as departments in St. Petersburg, Florida, Kinston, North Carolina, and Alliance, Nebraska, to deploy its own software, called TrustStat, to identify videos supervisors should review. “What we’re saying is, look, here’s an interaction which is statistically significant for both positive and negative reasons. A human being needs to look,” said Wender, the company’s founder.

TrustStat grew out of a project of the Defense Advanced Research Projects Agency, the research and development arm of the U.S. Defense Department, where Wender previously worked. It was called the Strategic Social Interaction Modules program, nicknamed “Good Stranger,” and it sought to understand how soldiers in potentially hostile environments, say a crowded market in Baghdad, could keep interactions with civilians from escalating. The program brought in law enforcement experts and collected a large database of videos. After it ended, Wender founded Polis Solutions, and used the “Good Stranger” video database to train the TrustStat software. TrustStat is entirely automated: Large language models analyze speech, and image processing algorithms identify physical movements and facial expressions captured on video.

At Washington State University’s Complex Social Interactions Lab, researchers use a combination of human reviewers and AI to analyze video. The lab began its work seven years ago, teaming up with the Pullman, Washington, police department. Like many departments, Pullman had adopted body cameras but lacked the personnel to examine what the video was capturing and train officers accordingly.

The lab has a team of around 50 reviewers — drawn from the university’s own students — who comb through video to track things like the race of officers and civilians, the time of day, and whether officers gave explanations for their actions, such as why they pulled someone over. The reviewers note when an officer uses force, if officers and civilians interrupt each other and whether an officer explains that the interaction is being recorded. They also note how agitated officers and civilians are at each point in the video.

Machine learning algorithms are then used to look for correlations between these features and the outcome of each police encounter.

“From that labeled data, you’re able to apply machine learning so that we’re able to get to predictions so we can start to isolate and figure out, well, when these kind of confluences of events happen, this actually minimizes the likelihood of this outcome,” said David Makin, who heads the lab and also serves on the Pullman Police Advisory Committee.

One lesson has come through: Interactions that don’t end in violence are more likely to start with officers explaining what is happening, not interrupting civilians, and making clear that cameras are rolling and the video is available to the public.

The lab, which does not charge clients, has examined more than 30,000 hours of footage and is working with 10 law enforcement agencies, though Makin said confidentiality agreements keep him from naming all of them.

Much of the data compiled by these analyses and the lessons learned from it remains confidential, with findings often bound up in nondisclosure agreements. This echoes the same problem with body camera video itself: Police departments continue to be the ones to decide how to use a technology originally meant to make their activities more transparent and hold them accountable for their actions.

Under pressure from police unions and department management, Tassone said, the vast majority of departments using Truleo are not willing to make public what the software is finding. One department using the software — Alameda, California — has allowed some findings to be publicly released. At the same time, at least two departments — Seattle and Vallejo, California — have canceled their Truleo contracts after backlash from police unions.

The Pullman Police Department cited Washington State University’s analysis of 4,600 hours of video to claim that officers do not use force more often, or at higher levels, when dealing with a minority suspect, but did not provide details on the study.

One lesson has come through: Interactions that don’t end in violence are more likely to start with officers explaining what is happening, not interrupting , and making clear that cameras are rolling.

At some police departments, including Philadelphia’s, policy expressly bars disciplining officers based on spot-check reviews of video. That policy was pushed for by the city’s police union, according to Hans Menos, the former head of the Police Advisory Committee, Philadelphia’s civilian oversight body. The Police Advisory Committee has called on the department to drop the restriction.

“We’re getting these cameras because we’ve heard the call to have more oversight,” Menos said in an interview. “However, we’re limiting how a supervisor can use them, which is worse than not even requiring them to use it.”

Philadelphia’s police department and police union did not respond to requests for comment.

Christopher J. Schneider, a professor at Canada’s Brandon University who studies the impact of emerging technology on social perceptions of police, said the lack of disclosure makes him skeptical that AI tools will fix the problems in modern policing.

Even if police departments buy the software and find problematic officers or patterns of behavior, those findings might be kept from the public just as many internal investigations are.

Because it’s confidential,” he said, “the public are not going to know which officers are bad or have been disciplined or not been disciplined.”

Umar Farooq was an Ancil Payne Fellow with ProPublica.