When police officers in Oakland, California, began wearing body cameras, the goal was clear: increase transparency. Triggered by national outrage over the Michael Brown filming in 2014the federal government, under President Obama, pledged $75 million to help law enforcement adopt the technology. The idea was to restore public trust, particularly in communities that felt targeted by police brutality. The cameras captured millions of hours of footage – of traffic stops, arrests and even casual conversations on street corners.

But the images collected remained largely unused. Beyond the courtroom and occasional media scrutiny, these clips sat in digital vaults, invisible to the public. As Benjamin Graham, a political scientist at the University of Southern California, points out: Scientific American“We spend a huge amount of money collecting and storing this data, but it’s almost never used for anything.”

Researchers like Graham are trying to change that. By reimagining body camera footage not just as evidence but as a vast repository of data, they aim to transform how policing operates. Using advanced artificial intelligence, these scientists leverage video transcripts to discover patterns in agent behavior. This approach, they say, could reshape police training and build trust in communities where that trust is shaken — and the results so far have been very promising.

Spotting trends in policing: from foul language to escalation

Jennifer Eberhardt, a psychologist at Stanford University, has spent years exploring police interactions. Working with Dan Jurafsky, a Stanford linguist, his team began analyzing body camera footage in 2014. The initial focus was on the Oakland Police Department, which had adopted the technology years earlier. For Eberhardt, the goal was to go beyond the surface and understand, in each moment, how the police communicate with citizens.

Eberhardt’s team developed an AI-based model to assess compliance with the language used by officers during traffic stops. They trained the system by analyzing nearly 1,000 transcripts and measuring how officers spoke to black and white drivers.

The findings were revealing: officers spoke with less respect to black drivers than to white drivers. They were less likely to explain the reason for the stop, provide reassurance, or express concern about the driver’s safety. These discrepancies existed regardless of the race of the officer, the reason for the stop, or the end result. When the Stanford team presented their results to the Oakland Police Department, they confirmed the long-held suspicions of the city’s minority communities.

These ideas weren’t just published in academic journals: they spurred action. Stanford researchers helped Oakland develop a “respect” module for their training programs, drawing directly on actual interactions recorded by the cameras. Officers were trained to use more respectful language and explain their actions to those they interacted with. The effects were tangible: after implementing the training, officers became more likely to reassure drivers and provide clear reasons for stops.

But the researchers didn’t stop there. In a separate study last year, they analyzed body camera footage to identify “linguistic signatures” that predicted when a traffic stop would turn into an arrest or search. As they scrutinized the first 45 words the officers spoke, they noticed a pattern. Apparently, giving direct orders without explanation often led to an unfavorable outcome for the driver. This pattern was disturbingly present in footage of George Floyd’s fatal encounter with police officers in 2020.

A data-driven path to reform or a political minefield?

Inspired by these findings, the Los Angeles and San Francisco police departments are now exploring similar projects. The Los Angeles Board of Police Commissioners, for example, tapped Graham’s team at USC to analyze 30,000 body camera videos, in an effort to understand the nuances of traffic stops. The San Francisco Police Department also partnered with Stanford researchers to evaluate a program focused on nonviolent communication.

However, the expansion of this approach faces significant obstacles. Departments may be reluctant to open their images to scrutiny, fearing what those scans might reveal. In some cases, reluctance stems from concerns about privacy or possible legal repercussions. But as Eberhardt points out, a systematic analysis of these interactions is crucial to understanding the real impact of police training. “A lot of this training that they’re getting now is just not rigorously evaluated,” she told SciAmerican. “We don’t know if what they’re learning… actually translates into real interactions with real people on the street.” »

“By undertaking these types of studies and making improvements in your department, it really helps to build trust in communities that have very low levels of trust,” added LeRonne Armstrong, former police chief of the Department of Oakland police in California.