Broadcast police communications may pose privacy risks, especially to Black men

The researchers found that event utterances contained the most references — about 60% — to gender, race/ethnicity, age and protected health information, which can be used to identify individuals. Nearly 68% of utterances that included a sociodemographic indicator used male gendered terms, and approximately 69% of those utterances referred to Black people, according to the researchers.

“Our findings contribute to a larger body of evidence about racial disparities in policing. What is novel here is the data source: radio transmissions,” Graziul said. “Despite prolific use by police systems around the world, few have explored what this means of communication can tell us about how policing operates in practice. Disproportionate mentions of Black people reflect a novel way to observe how officers’ attention is unevenly distributed across racial/ethnic groups, and identifying this disparity helps us understand challenges to the ethical use of this data source for research, like preventing leakage of sensitive personal information, which would impact Black communities substantially more than other communities.”

After examining the utterances, the team tested a large language model (LLM), a widely used artificial intelligence tool, to determine its capacity to find personal information in the transcripts. Despite the unique nature of the language involved with broadcast police communications, the LLM detected personally identifiable information with high accuracy, highlighting the risk of privacy vulnerability. Bad actors, such as identity thieves, could use AI technology to quickly find and misuse the personal information in transcripts of police radio activity, according to the researchers.

“This work reveals a concerning trend of racial inequality in terms of the exposure of sensitive information during police radio transmissions,” said Pranav Narayanan Venkit, graduate student pursuing a doctoral degree in informatics in the College of Information Sciences and Technology and first author on the paper. “This study may help researchers and developers give more thought to interactions between LLM and different segments of society — the policing community, minority populations and various other populations — to identify biases and protect personal information.”

Miranda Goodman, who graduated with her bachelor’s degree from Penn State this past summer, and fourth-year Penn State student Samantha Kenny also contributed to this work.

The National Institute of Minority Health and Health Disparities of the National Institutes of Health supported this work. This research was also supported by the Urban Resiliency Initiative, which is led by Margaret Beale Spencer, distinguished professor emerita at the University of Chicago and a principal investigator of this larger project.