AI analysis of police body camera video reveals what typically happens during traffic stops

A decade ago, then-President Barack Obama proposed spending $75 million over three years to help states buy police body cameras to expand their use. The move came after the killing of teenager Michael Brown, for which no body camera footage existed, and was designed to increase transparency and build trust between police and the people they serve.

Because the first funds were allocated in 2015, tens of millions of traffic stops and crashes, street stops, arrests, and more were recorded by these small digital devices that police officers attach to their uniforms or winter jackets. The video has been considered useful as evidence in controversial incidents such as the one that led to the death of George Floyd in Minneapolis in 2020. Using cameras can also deters bad behavior police officers in their interaction with the public.

But unless something tragic happens, body camera footage usually goes unnoticed. “We spend so much money collecting and storing this data, but it’s almost never used for anything,” says Benjamin Graham, a political scientist at the University of Southern California.

About supporting scientific journalism

If you like this article, please consider supporting our award-winning journalism subscription By purchasing a subscription, you help ensure the future of the amazing stories of discovery and ideas that shape our world today.

Graham is one of the few scholars who reinterprets this footage as data, not just evidence. Their work uses advances in natural language processing, which relies on artificial intelligence, to automate the analysis of video recordings of interactions between citizens and police officers. The resulting data allowed police departments to identify problems with the police, find ways to fix them and determine whether the corrections improve the behavior.

Only a small number of police agencies have so far opened their databases to researchers. But if that footage were routinely analyzed, it would “really make a difference,” says Jennifer Eberhardt, a psychologist at Stanford University who pioneered this line of research. “We can see beat by beat, moment by moment, how the interaction unfolds.”

In papers published over the past seven years, Eberhardt and her colleagues have examined body camera recordings to find out how the police talk differently to white people and black people and what type of conversation can either gain a person’s trust or portend an undesirable outcome, such as handcuffs or arrest. The results improved and improved police training. In a study published in PNAS Nexus in September researchers showed that the new training changed the behavior of the officers.

“Having this kind of research and improvement in your department helps build trust in communities that have very low levels of trust,” says LeRon Armstrong, former police chief of the Oakland, California Department of Police, who has a long history of working with the Stanford team.

The approach is slowly learned. Inspired by Stanford University’s findings, the Los Angeles Board of Police Commissioners, which oversees the Los Angeles Police Department (LAPD), asked USC to help dig into the department’s records. A project is currently underway to analyze 30,000 body camera videos spanning a year of traffic stops. The Stanford group is also working with the San Francisco Police Department to use body camera footage to evaluate a program in which its officers travel to Birmingham, Alabama, to learn about the civil rights movement and nonviolence principles.

Stanford’s work began in 2014 after a scandal involving the Oakland Police Department. Four Oakland, California police officers, known as the “Riders,” were accused of roughing and drugging innocent people, among other crimes, back in the late 1990s. Of the 119 plaintiffs, 118 were black. So as part of the $10.9 million settlement, the department had to collect data on vehicle and pedestrian stops and analyze them by race. More than a decade after the deal was reached, the department’s federal monitor turned to Eberhardt for help.

The plaintiffs’ attorneys told Eberhardt that what they most wanted to know was what happened next the lights of the cruiser came on— why the police stopped people and how the interaction took place. The department was the first to introduce body cameras, which it put into operation about five years ago. “You actually have the footage,” Eberhardt recalls telling them, though no one in the department thought to use it for that purpose.

Eberhardt enlisted Dan Jurafsky, a Stanford linguist and computer scientist, and his then-student Rob Voigt, now a computational linguist at Northwestern University, to develop an automated way to analyze the video transcripts for nearly 1,000 stops. The researchers set out to determine whether police officers speak less respectfully to black drivers than to white drivers. They first got people to appreciate the respectfulness of excerpts from the transcripts. They then built a computational model that linked the ratings to different words or phrases and gave those statements a numerical weight. Expressing concern for a driver, for example, was rated as highly respectful, while addressing him by name was less respectful.

The model then scored all of the officers’ speech over a month of stops, and the researchers linked those scores to the race of the person stopped, among other variables. They found the clear racial inequality in the respectful language of the officers. When interacting with black drivers, police officers were less likely to give a reason for the stop, give reassurances, or express concerns about the driver’s safety, for example. The respect gap existed throughout the interaction and was independent of the officer’s race, the reason for the stop, its location, or its outcome.

These first results, published in 2017, had a profound impact on Auckland. “When Stanford released the results, it was almost like a sigh of relief for minority communities,” Armstrong says. “It confirmed a concern that people have always had and has forced the department to reevaluate how we train our officers to interact with our community.”

The Stanford team used the findings to develop a “respect” module for the department’s procedural justice curriculum. Procedural justice seeks to build fairness into police procedures. In addition to emphasizing respect, this may involve police officers explaining their actions to others and giving those individuals a chance to express their point of view. As part of this effort, the team used their computational model to identify real-world interactions that were particularly respectful and disrespectful. “As a learning example, it feels much more legitimate to the learner” than fictional scenarios, Jurafsky says. “(The officers) recognize their language.”

After the training took effect, the researchers conducted another body camera study to determine if the officers were using what they had learned. The Stanford University team compared key features of the officers’ language during 313 stops that occurred four weeks before the training with those made during 302 stops during the four weeks after the training. Researchers found that police officers who received the training were more likely to express concerns about the safety of drivers, offer reassurance and explain clear reasons for stops, they reported in a September report PNAS Nexus study.

Systematic analysis of body camera footage, Eberhardt says, offers a promising way to understand what types of police training are effective. “A lot of the training they do now is just not rigorously evaluated,” she says. “We don’t know if all the things they’re learning in these trainings … actually translate to real interactions with real people on the street.”

In a study published last year, researchers at Stanford University analyzed body camera recordings to find language related to “escalation result” to stop traffic, such as handcuffing, searching or arresting. Using footage from 577 stops of black drivers in an unknown city, they found what Eberhardt calls a “linguistic signature” for escalation in the first 45 words spoken by the officer: ordering the driver from the start and not giving a reason for the STOP. “The combination of those two factors was a good signal that the stop would result in the driver being handcuffed, searched or arrested,” she says.

None of the stops in the study involved the use of force. But the researchers wondered if the signature they found would be present in the footage of the police interaction that led to Floyd’s death. It was. In the first 27 seconds of the encounter (about as long as it takes a police officer to say 45 words during a traffic stop), the officer gave only orders and did not tell Floyd why he was stopped.

The USC team recruited a diverse group of people, including formerly incarcerated and retired police officers, to evaluate interactions captured by LAPD body cameras for civility, respect and other aspects of procedural justice. The team plans to use AI advances to capture these perspectives in ways that, for example, reveal why a statement intended to be funny or respectful can be perceived as sarcastic or disrespectful. “The biggest hope is that our work can improve the training of Los Angeles police officers, have a data-driven way to update and change training procedures to better match the populations they serve,” said USC cognitive scientist Morteza Deghani. , who jointly manages the project with Graham.

Politics can prevent police departments from sharing footage with academics. In some cases, departments may be reluctant to highlight systemic issues. However, in the future, departments may independently analyze the footage. Some private firms, for example TRULEO and Polis Solutions— already offer software for this purpose.

“We’re getting closer to departments being able to use these tools rather than just training exercises,” says Nicholas Kemp, a social psychologist at the University of Michigan who worked on Eberhardt’s team. But commercial models tend not to be completely transparent—users can’t inspect their components—so some scientists, including Kemp and Degany, are wary of their results.

The USC team plans to make the verifiable language models it created available to the LAPD and other police departments so they can regularly monitor officers’ interactions with the public. “We need to have much more detailed information about how these day-to-day interactions happen. That’s a big part of democratic governance,” Graham says.