# Session 2: Leveraging Technology to Improve Police-Community Relations Source: https://www.youtube.com/watch?v=OiKp49mp4GU Recap page: https://rapidrecap.app/video/OiKp49mp4GU Generated: 2025-12-11T18:41:07.096+00:00 --- ## Quick Overview The research team successfully leveraged technology, including Large Language Models (LLMs) and human annotators, to analyze 1.3 million pieces of body-worn camera footage from two Bay Area police departments over three years, revealing systemic racial disparities in officer language, such as officers being significantly more likely to issue orders and less likely to state the reason for a stop to Black drivers compared to White drivers. **Key Points:** - The project analyzed 1.3 million pieces of body-worn camera footage from two Bay Area police departments over three years, involving 2,600 officers and 200k administrative records. - LLMs achieved 71% accuracy in predicting whether a stop would escalate based only on the officer's first words. - Officers were significantly less likely to state the reason for the stop and more likely to issue an order in their first 45 words when addressing Black drivers compared to White drivers. - Black drivers received measurably less respectful language (a negative score of -0.1) compared to White drivers in the high-redlined Area 1 & 5, while stops in lower-redlined areas (2, 3, & 4) showed less disparity (-0.06). - Survey results indicated Black Americans (34.94%) rated their treatment as very bad/very well compared to White Americans (51.64%) in high-redlined areas. - The team developed new tools to analyze longitudinal data, allowing for the evaluation of training effectiveness on officer behavior in real-world interactions, which was previously impossible at scale. - The research concluded that police-community relations are impacted by systemic issues, and technology can be used to foster trust and accountability by identifying these patterns. ![Screenshot at 04:05: The slide explicitly stating the core finding: 'policing eq trust', setting the stage for the research goal of finding technology-based solutions to bridge this trust gap.](https://ss.rapidrecap.app/screens/OiKp49mp4GU/00-04-05.png) **Context:** This presentation, titled "Leveraging Technology to Improve Police-Community Relations," was delivered by a multidisciplinary team including Jennifer Eberhardt (Social Psychologist), Dan Jurafsky (Linguist/Computer Scientist), Ralph Banks (Law Professor), and Benoît Monin (Psychology Professor). The project focused on using AI and human annotation to systematically analyze thousands of police-citizen interactions captured on body-worn cameras, aiming to uncover subtle linguistic and behavioral patterns that contribute to disparities in police-community relations, particularly in the context of historical housing segregation (redlining). ## Detailed Analysis The presentation detailed a project leveraging technology, specifically Large Language Models (LLMs) and human annotation, to analyze 1.3 million pieces of body-worn camera footage from two Bay Area police departments over three years. The core finding is that racial disparities exist in officer language and behavior during traffic stops. The team used LLMs to predict stop escalation based on the officer's first words with 71% accuracy. Analyzing the initial 45 words, officers were less likely to state the reason for the stop and more likely to issue an order when addressing Black drivers compared to White drivers. Further analysis of respect scores showed a greater negative impact (-0.1 vs -0.06) in historically redlined areas (Areas 1 & 5) for Black drivers compared to less-redlined areas (Areas 2, 3, & 4). Survey data corroborated this, showing Black Americans in redlined areas rated their treatment as very bad significantly more often (34.94%) than White Americans (51.64% rated treatment very well, implying a large gap in perceived negative treatment). The team developed a processing pipeline involving professional transcription and expert labeling to create a canonical form of the data, enabling the evaluation of training effectiveness and policy changes in real-time, moving beyond anecdotal evidence to systematic analysis of police-public interactions. ### Project Introduction and Team - The project, titled 'Leveraging Technology to Improve Police-Community Relations,' involved main PIs Jennifer Eberhardt and Ralph Banks, and Co-PIs Dan Jurafsky and Benoît Monin. - The goal was to transform analysis of police-community interactions using AI and LLMs, referencing the George Floyd footage as a catalyst for this work. - The project analyzed 1.3 million pieces of footage over three years from two Bay Area departments. ### Linguistic Analysis of Disparity - The analysis of officer language showed racial disparities in respect, with officers using less respectful language toward Black drivers than White drivers, especially in areas with historical redlining. - Early analysis of officer's first 45 words showed they were less likely to state the reason for the stop and more likely to issue an order to Black drivers. ### Data Scale and Methodology - The project utilized 1.3 million videos, 200k admin records, and 550 terabytes of data, requiring professional, vetted human transcribers and expert labelers. - The team created a canonical form to harmonize data from different departments for analysis. ### Impact of Redlining - A map overlaying redlining data with stop outcomes showed that stops in historically redlined areas (Areas 1 & 5) resulted in higher rates of handcuffing (41% vs 27%) and searching (39% vs 23%) for Black drivers compared to less-redlined areas (Areas 2, 3, & 4). - Black drivers in redlined areas reported significantly worse treatment (34.94% rated 'Very Bad - Very Well') compared to White Americans (51.64% rated 'Very Well'). ### Measuring Cooperation and Future Work - The team is developing tools to measure driver cooperation and assess the impact of police training interventions in real-time, moving beyond retrospective analysis of isolated incidents to systemic evaluation. ### Panel Discussion - The panel included the main PIs (Jennifer Eberhardt, Dan Jurafsky, Ralph Banks, Benoît Monin) and other team members (Mikaela Spruill, Julia Proshan, etc.) discussing the research and its implications for policy and trust. ![Screenshot at 00:00: Title slide listing the main and co-Principal Investigators for the project.](https://ss.rapidrecap.app/screens/OiKp49mp4GU/00-00-00.png) ![Screenshot at 01:50: A visual representation of the George Floyd body-camera footage analysis, showing dialogue attribution to 'Officer Lane' and 'George Floyd' categorized by speech function \(Exclamation, Apology, Plea, Proclaim Innocence, Request for reason, Explanation\).](https://ss.rapidrecap.app/screens/OiKp49mp4GU/00-01-50.png) ![Screenshot at 04:05: Slide summarizing the core message: 'policing eq trust', highlighting the need for technological solutions to bridge the trust deficit.](https://ss.rapidrecap.app/screens/OiKp49mp4GU/00-04-05.png) ![Screenshot at 08:57: Slide detailing the challenges in scaling the work, noting the need for professional, fingerprinted transcribers and highly-trained expert labelers because discerning the 'reason' is subtle.](https://ss.rapidrecap.app/screens/OiKp49mp4GU/00-08-57.png) ![Screenshot at 34:04: Slide illustrating that reasons for stops can be categorized as 'Easy' \(stating a clear violation\) or 'Hard' \(requiring justification or reference to policy\).](https://ss.rapidrecap.app/screens/OiKp49mp4GU/00-34-04.png)