Members of a University of Southern California program are developing an AI program meant to help law enforcement agencies improve their interactions with the communities they serve.
The Everyday Respect Project partnered with the Los Angeles Police Department to analyze body camera footage of 1,000 random traffic stops. Now, they are using what they have found to train an AI model to look through countless hours of videos for critical elements of good policing — respect and de-escalation.
Benjamin Graham is an associate professor of political science at USC and helps to manage the project, which is being conducted by a team of professors, students and members of the community.
”LAPD conducts, give or take, a thousand stops a day,” Graham told LAist. He said those stops lead to thousands upon thousands of hours of body camera footage.
In most cases, Graham said, the body camera videos are uploaded to the cloud and never seen. He said only an automated program could sort through this rich source of data, analyze it and reveal the stories it holds.
With AI able to look through this data, Graham said police and sheriffs departments across the country could identify officers who are best able to communicate respectfully during traffic stops to be given promotions or training positions. They also could find the best techniques for officers to bring down the temperature in high-stress situations.
Georgetown University, the University of California, Riverside, and the University of Texas at Austin also are working with USC on the project.
What they’ve found so far
Graham told LAist researchers spoke with a wide range of community stakeholders, including community organizations that are critical of law enforcement and working police officers, to understand different perspectives of what separates a good traffic stop from a bad one.
They heard from thousands of Angelenos through surveys and interviews, reviewed LAPD training materials and rode along with officers on the streets. Graham said they focused on those diverse community perspectives throughout the project.
“ We have former law enforcement officers who are annotating this data,” he told LAist. “We have individuals who have been arrested before, and we have a lot of Angelenos from ... a range of ages, races, genders, professional backgrounds.”
Graham presented some of the Everyday Respect Project’s findings to the L.A. Board of Police Commissioners on Dec. 16.
He said researchers involved in the project analyzed 500 traffic stops in which LAPD officers conducted searches of the drivers they pulled over and another 500 stops in which there were no searches.
They found in cases when no search was conducted, some drivers were treated differently by officers based on their perceived race and wealth.
Researchers found that Black drivers were treated with more respect than Hispanic drivers. White drivers were pulled over least often, and the researchers did not find a significant difference between how white and non-white drivers were treated.
Of drivers who were stopped by police but weren’t searched, the researchers found those who were perceived as more wealthy also were treated with more respect by officers.
They did not find significant differences in how drivers were treated due to perceived race or wealth in stops in which searches were conducted.
Across all stops, the researchers found the more respect they perceived an officer showing to a driver they pulled over, the more legitimate the researchers would tend to rate the stop overall.
The LAPD has not responded to LAist’s request to comment on these findings.
Training AI to tell good traffic stops from bad
After analyzing and manually taking detailed notes on the first 1,000 traffic stops, Graham said the researchers are using what they’ve found to build an AI tool that can do the same thing — but is able to cover vastly more data and is accessible free of charge for any law enforcement agency.
To do this, Graham said team members use their notes as training data for the AI model.
By having humans label a number of things that happened or didn’t happen in videos of traffic stops, Graham told LAist, the AI model they are developing can learn to predict what humans will say about other videos.
“You're trying to train a model to do the same job that a human being does when it watches the video,” he said.
These notes include things like whether a search happened and at what time, whether officers explained the reason for the stop, if the driver complied with requests from the officer and any efforts by officers to de-escalate tense situations.
Graham said other companies have been working on AI tools to sell to departments, as well, but that the Everyday Respect Project is unique in its effort to build community perspectives into the program that will be fully open-source and open-science.
That means anyone can see exactly how the program works and the research behind it.
What could be coming next?
Graham said the Everyday Respect Project will be working through the winter and spring to improve the AI model and use it to analyze more LAPD bodycam videos. Then they will present their new findings to the police commissioners and release their AI model to the public.
It still is uncertain whether LAPD will use the Everyday Respect Project’s AI program once it is completed, but Captain Shannon White of LAPD’s Strategic Planning and Policies Division told the police commission Dec. 16 that the department looks forward to using the group’s research to spark “actionable change within the department.”
The LAPD has not responded to LAist’s questions about whether or how it will use the AI program once it is made available.
How to reach me
If you have a tip, you can reach me on Signal. My username is jrynning.56.
Graham told LAist other departments have shown interest in continuing to work with the Everyday Respect Project and the program in the future.
He said they may soon be working with the Rochester Police Department in New York on a trial to find the most effective de-escalation techniques for officers.
“ That's an incredible piece of learning that we can bring to improve policing, to improve officer safety, community safety, the whole nine yards,” Graham told LAist.