Three former Memphis police officers have just been convicted of obstruction of justice and witness tampering in the January 2023 beating death of Tyre Nichols. Two others have already pleaded guilty.

Graphic police body camera video documented the incident after Nichols was stopped for allegedly trying to beat a red light. Officers swarmed his car with guns raised, shouting expletive-laden orders for him to get out and show his hands. “I didn’t do anything,” Nichols, a FedEx worker, protested. “Get out of the f**** car,” an officer yelled.

At no point did the officers explain to Nichols why he was pulled over. Nichols was handcuffed and brutally beaten. He died three days later.

The footage, like other video of police encounters, quickly went viral and was key evidence in the officers’ federal trial.

Once met with resistance, body cameras are now recognized as potential tools for reform – not as evidence, but as data.

Stanford researchers are at the forefront of using AI to analyze bodycam footage to improve policing.

“We can use technology to address pressing social problems,” Stanford social psychologist Jennifer Eberhardt told experts in law enforcement, policy, and technology during a Summit on AI, Body-Worn Cameras, and the Future of Policing hosted at Stanford on Sept. 12.

“The issues are complex and challenging,” Eberhardt said, “yet we are compelled to work on them given the huge possibilities for advancement.”

Sponsored by Stanford’s Business, Government, and Society Initiative, the day-long series of working sessions explored how body camera footage, combined with advances in AI, could reshape law enforcement, offering new ways to analyze police-civilian interactions and improve public safety while addressing critical issues of privacy and civil liberties. Attendees included police chiefs, top officials in the California Department of Justice, other academic researchers, and experts in issues of privacy and public access to government data.

Body-worn cameras are the “biggest technology investment in policing” in a generation, said Eberhardt, the William R. Kimball Professor at the Graduate School of Business.

And yet, much of their footage is underutilized.

“What about the thousands and thousands of hours of routine footage?” Eberhardt said. “The vast majority is never examined. Eventually it’s deleted, and that’s a lost opportunity.”

If law enforcement agencies aided by researchers can leverage this trove of untapped data, Eberhardt said, it could be incredibly beneficial to public safety.

Toward this end, Eberhardt’s team at Stanford SPARQ – short for Social Psychological Answers to Real-World Questions – is collaborating with Dan Jurafsky’s Natural Language Processing Lab to analyze bodycam footage and identify “linguistic signatures” that predict whether an encounter will escalate to handcuffs, arrest, or violence.

“I believe research is key to our future of policing,” said San Francisco Police Chief Bill Scott, who recalled early resistance and skepticism toward body cameras. “Now there is appreciation.”

Scott reflected on the impact of the Rodney King case in 1991, when the videotaped beating of King by Los Angeles police officers shocked the world. That marked a turning point for policing, Scott said, and raised the question, “Should cameras be a gotcha tool or something more holistic?”

“We’d made it into a gotcha,” Scott said. “It was accountability, accountability, accountability, and punishment, punishment, punishment.”

As law enforcement continues to evolve, new challenges emerge. Venus Johnson, California chief deputy attorney general, emphasized that while progress has been made, many questions remain. She pointed to issues such as data storage, usage, and funding, as well as the complexities of when to activate or deactivate cameras, particularly in sensitive situations involving crime victims.

While there was “a lot of hope” when body cameras were introduced, “it’s inconclusive whether body-worn cameras have reduced the use of force,” said Max Isaacs, director of technology law and policy at New York University’s Policing Project. AI analysis offers the potential for deeper insights, Isaacs added.

Eberhardt noted that when she began her research on police interactions in 2014, her team transcribed footage by hand. She was asked to analyze Oakland Police Department’s car stop data to determine whether there were significant racial disparities and suggest ways to improve as part of a consent decree related to a lawsuit involving four rogue Oakland police officers known as the “Riders.”

As the volume of data has grown, Eberhardt’s team has turned to AI, collaborating with Jurafsky’s lab to run computerized language analyses on the recordings.

“Ten years ago, Dan Jurafsky and I established an interdisciplinary team of researchers here at Stanford to work on these issues,” Eberhardt said. “Rather than using police body-worn camera footage as evidence – in one police encounter at a time – we treat the footage as data, analyzing thousands of encounters at once.”

One of their key findings: An officer’s first words can significantly impact the outcome of a traffic stop, which is the most common interaction between police and the public, with more than 18 million drivers pulled over each year.

“In fact, the first 45 words an officer utters – roughly the first 27 seconds of a stop – predict whether that stop will end with the officer handcuffing, searching, or arresting the driver,” Eberhardt told the summit attendees.

Moreover, Black male drivers, like Nichols, are not only more likely to be stopped but also more likely to be searched, handcuffed, and arrested.

“The Tyre Nichols’ stop is an excellent example of what we found – that is, in the first moments of the stop, the officers give orders, but do not give the driver a reason for the stop,” Eberhardt told Stanford Report. “These are the two elements – what we call the ‘linguistic signature’ – of an escalated outcome, where the driver is handcuffed, searched, or arrested by the end of the stop.”

The sheer volume of data generated by body cameras is overwhelming for law enforcement.

“It is tedious to review recordings,” Scott said. “Technology is absolutely necessary.”

Some law enforcement agencies have contracted with a private company, Truleo, which uses AI to transcribe and analyze footage, but this raises concerns about privacy, data ownership, and security.

These issues were central to the summit’s collaborative workshops, led by Dan Sutton, director of justice and safety for the Stanford Center for Racial Justice, who previously designed police reform initiatives for Rhode Island.

“In the workshop sessions we were able to bring together people who rarely have the opportunity to collaborate, like police chiefs, AI experts, and privacy advocates,” Sutton said. “This approach allowed us to look at challenges from multiple angles and generated some new ideas and potential pathways that we’re eager to explore.”

Reflecting on the overall summit experience, Eberhardt called the gathering “galvanizing.”

“We came away with greater appreciation for this opportunity that is before us – the opportunity to leverage technology to improve police-community relations … and to do so not one interaction at a time, not one police department at a time, but across the industry,” she said.