Culture & Society

Will a Police Stop End in Arrest? Listen to Its First 27 Seconds.

Researchers have identified a linguistic signature that can predict whether encounters with cops will escalate. Black drivers hear this pattern as well.

May 30, 2023

| by Sara Harrison
A photo of a Black woman in a car, looking nervously back at the flashing lights of the police car that has just pulled her over. Credit: iStock/RichLegg

A police officer’s first 45 words hold clues as to how a stop will unfold. | iStock/RichLegg

When Minneapolis police officer Derek Chauvin killed George Floyd in May 2020, he was surrounded by cameras. Bystanders took videos on their cell phones, security cameras were rolling at the convenience store where Floyd had just purchased cigarettes, and Chauvin’s fellow officers’ bodycams recorded the scene.

In the following months, that footage was replayed and reviewed by journalists, policing experts, attorneys, and activists who wanted to know how a stop involving the alleged use of a counterfeit $20 bill turned deadly so quickly. Most focused on Floyd’s desperate final minutes. Yet the clues that this incident could become violent were evident in the opening seconds of a police officer approaching Floyd’s car, well before Chauvin arrived.

Every year, police officers in the United States stop nearly 19 million drivers. In about 2% of those stops, officers use force. Few are analyzed as intensely as Floyd’s murder, but these exchanges contain telling details about cops’ treatment of members of the public, particularly Black men.

“Even during what many would consider routine encounters, where no force is used, Black drivers are often treated differently from white drivers, and that differential treatment can erode their trust in police,” says Jennifer L. Eberhardt, a professor of organizational behavior at Stanford Graduate School of Business and professor of psychology at Stanford University. “Black drivers worry about escalation — whether they will be handcuffed, searched, or arrested — despite the fact that they are much more likely than white drivers to be stopped for relatively minor violations that pose little threat to public safety.”

In a paper recently published in the Proceedings of the National Academy of Science, Eberhardt and her coauthors reveal new clues about how some police stops escalate. By analyzing the first words spoken by an officer stopping a car, her team was able to predict whether the driver would be searched, handcuffed, or arrested.

“There’s a linguistic signature to escalated stops,” Eberhardt says. She and her coauthors also found that Black men are attuned to this linguistic signature.

What the First 45 Words Say

Eberhardt, a social psychologist, studies how racial bias can influence perceptions and interactions in a variety of environments, from classrooms to courtrooms. Her previous research has demonstrated that police officers are more likely to identify African American faces as criminal and that defendants who appeared “more stereotypically Black” were more likely to receive a death sentence in cases involving white victims.

Rather than using footage as evidence in a particular case, we’re trying to treat footage as data.
Jennifer L. Eberhardt

Since 2014, Eberhardt has worked with Dan Jurafsky, a professor of linguistics and computer science at Stanford, to analyze audio from police officers’ body-worn cameras. “Rather than using footage as evidence in a particular case, we’re trying to treat footage as data,” she says. By analyzing thousands of car stops captured by cameras, Jurafsky and Eberhardt have begun to uncover patterns that reveal disparities in how police treat community members. In 2017, their interdisciplinary team found that police officers consistently address Black drivers with less respect than white drivers during traffic stops. Their ongoing investigation of policing is in collaboration with Stanford SPARQ, a behavioral science “do tank” that Eberhardt codirects.

For the current paper, a team led by former Stanford postdoctoral scholar Eugenia Rho (now an assistant professor of computer science at Virginia Tech) analyzed nearly 600 car stops recorded by police bodycams over one month in a mid-sized American city. Because Black people are disproportionately likely to be pulled over by the police, the researchers focused only on stops involving Black drivers. During the period the researchers examined, around 15% of Black drivers were searched, handcuffed, and/or arrested. (In contrast, less than 1% of white drivers pulled over during this time experienced one of these escalated outcomes.)

To understand the role of the officer’s initial words, the team fed the first 45 words spoken to the driver in each stop to a series of natural language processing (NLP) models. “Modern large language models are very good predictive tools,” says Jurafsky. Looking at just the first 45 words of an interaction (roughly the first 27 seconds), their best-performing large language model could predict whether a stop would lead to a search, handcuffing, or arrest with over 70% accuracy.

The team then zoomed in to try to understand what was happening in the language used by the officers. “Interpretative AI tools can help understand what makes the language model make its predictions, but it’s crucial to also do a careful linguistic analysis to understand the officer language itself,” Jurafsky says. The team had research assistants read the officers’ opening words and sort them into six categories, such as explaining the reason for the stop, asking for documents, or giving a command.

Across hundreds of stops conducted by 200 officers, the researchers were able to discern a systematic pattern that distinguishes uneventful stops from escalated stops. The team found that in these escalated stops involving Black drivers, cops were significantly more likely to issue orders and less likely to explain why they pulled the driver over.

Recording a Dangerous Dynamic

Next, the researchers wanted to see if Black men were also attuned to these patterns. They asked 188 Black male participants to listen to the opening 45 words of a series of car stops. The researchers then asked participants how they would feel if they were the driver in that situation. Participants rated the officer’s demeanor, including whether they were aggressive or friendly, condescending or respectful. They were also asked how confident they were that the stop would end with the driver being searched, handcuffed, or arrested. Finally, the researchers asked participants how worried they would be that this interaction might end with police officers using or threatening to use force against the driver.

On average, the Black men in the study reported feeling more negative emotions after listening to only the opening seconds of clips from stops that escalated. They also rated the officers’ demeanor in those clips more negatively and expressed greater confidence that the stop would end with the driver being searched, handcuffed, or arrested. While none of the stops in the study ended with force, participants were prepared for the possibility of violence. When listening to the escalated stops, they said they felt more worried about the officer using force.

“The problem is there at the beginning,” Jurafsky says. This isn’t meant to suggest that the outcome of a stop is predetermined by the first words that come out of an officer’s mouth. But these studies show that even the first 5% of an interaction can signal a dangerous trajectory — and Black men are picking up on those cues. As the authors write, “An officer’s speech, then, can set into motion negative perceptions and emotions of the driver — sparking a dynamic that erodes trust and undermines the relationship between police and those they are meant to serve.”

That dynamic was apparent in the minutes leading up to George Floyd’s murder. The officer who initially approached Floyd’s car issued 9 orders in 27 seconds and provided no explanation of why he was stopping Floyd. As Floyd responded with apologies, pleas, and expressions of innocence, Officer Thomas Lane issued increasingly aggressive orders that he keep his hands on the steering wheel.

Currently, most police bodycam footage goes unwatched — a missed opportunity to learn from the majority of stops and calls that don’t result in arrests or the use of force. Eberhardt and Jurafsky hope their analysis could be used to design trainings that change how these kinds of encounters unfold. They are also looking into ways police departments could use NLP to evaluate and improve their officers’ relationships with the communities they patrol.

“We believe we could leverage the power of the camera to not only capture police encounters as they unfold, but to change the course of them,” Eberhardt says. “This is a goal that is well within our reach.”

The study was coauthored by Eugenia H. Rho, assistant professor of computer science at Virginia Tech; Maggie Harrington, PhD student in psychology at Stanford University; Yuyang Zhong, a computational social science researcher at University of Michigan; Reid Pryzant, former PhD student in computer science at Stanford; Nicholas P. Camp, assistant professor of organizational studies at University of Michigan; Dan Jurafsky, professor of linguistics in the School of Humanities and Sciences and of computer science in the School of Engineering at Stanford; and Jennifer L. Eberhardt, professor of organizational behavior at Stanford Graduate School of Business.

For media inquiries, visit the Newsroom.

Explore More