# Session 2: Leveraging Technology to Improve Police-Community Relations

Source: https://www.youtube.com/watch?v=fZC2eX-RGts
Recap page: https://rapidrecap.app/video/fZC2eX-RGts
Generated: 2025-10-30T16:39:42.807+00:00

---
## Quick Overview

The research team leveraged large language models (LLMs) and a custom data infrastructure built over three years from 1.3 million body-worn camera videos across two police departments to analyze racial disparities in officer-driver interactions, finding that officers use less respectful language and are more likely to issue orders to Black drivers compared to White drivers, even before the drivers speak, but that the data-driven approach can also be used to evaluate and potentially improve police training and community relations.

**Key Points:**
- The project analyzed 1.3 million pieces of body-worn camera footage over three years from two Bay Area police departments involving 2,600 officers.
- The analysis revealed significant racial disparities: officers spoke less respectfully and were more likely to issue orders to Black drivers than White drivers.
- Black drivers were found to be handcuffed, searched, and arrested at higher rates across interactions in historically redlined areas (Areas 1 & 5) compared to non-redlined areas (Areas 2, 3, & 4).
- LLMs achieved 71% accuracy in predicting if a stop would escalate based only on the officer's first words.
- The research team developed a data infrastructure allowing for real-time analysis of officer-public interactions, which is crucial for evaluating training effectiveness.
- Explicit data analysis showed that Black drivers received fewer justifications for stops (e.g., 'Running a light is dangerous') and more direct orders compared to White drivers.

![Screenshot at 00:05: An introductory slide displaying the main presenters: Jennifer Eberhardt, Ralph Banks, Dan Jurafsky, and Benoît Monin, setting the context for the presentation on leveraging technology to improve police-community relations.](https://ss.rapidrecap.app/screens/fZC2eX-RGts/00-00-05.png)

**Context:** This presentation details a research project titled "Leveraging Technology to Improve Police-Community Relations," led by Jennifer Eberhardt (Main PI) and Co-PIs Dan Jurafsky and Benoît Monin, along with Ralph Banks. The core of the work involved analyzing massive amounts of body-worn camera footage to quantitatively study racial disparities in police-community interactions, specifically focusing on the language used by officers during traffic stops. The project highlights the potential of AI and large language models to analyze complex human interactions at scale to promote accountability and systemic change.

## Detailed Analysis

The presentation introduces a project focused on using technology, specifically Large Language Models (LLMs) and data analysis, to improve police-community relations. The research team, led by Jennifer Eberhardt and Dan Jurafsky, analyzed 1.3 million pieces of body-worn camera footage collected over three years from two Bay Area police departments involving 2,600 officers. Key findings demonstrated racial disparities in officer language: officers used less respectful language and were more likely to issue direct orders to Black drivers than White drivers, starting within the first 45 words of the stop, even before the driver spoke. The analysis also showed that stops occurring in historically redlined areas (Areas 1 & 5) resulted in higher rates of handcuffing, searching, and arrest for Black drivers compared to those in non-redlined areas (Areas 2, 3, & 4). The researchers developed a scalable data infrastructure and AI tools (like BERT) to automate the analysis of transcripts, achieving 71% accuracy in predicting stop escalation based on the officer's opening words. The presentation concluded by discussing how this technology can be used proactively to evaluate training and foster accountability, rather than just analyzing past incidents.

### Project Overview and Leadership

- Main PI Jennifer Eberhardt and Co-PIs Dan Jurafsky, Ralph Banks, and Benoît Monin led the project on improving police-community relations through technology
- Project utilized 1.3 million body camera videos from two departments over three years involving 2,600 officers
- The work aims to build trust and accountability in policing.

### Racial Disparities in Officer Language

- Analysis showed police officers are more polite to White drivers than Black drivers in terms of respect in their tone of voice and language used
- Black drivers were more likely to receive direct orders, while White drivers received more reassurance and information.

### Redlining and Stop Outcomes

- Stops in historically redlined areas (Areas 1 & 5) resulted in significantly higher rates of handcuffing (41% vs 27%), searching (39% vs 23%), and arrest (11% vs 8%) for Black drivers compared to non-redlined areas (Areas 2, 3, & 4).

### AI and LLM Application

- LLMs (like BERT) achieved 71% accuracy in predicting if a stop would escalate based only on the officer's first words
- Automated analysis allows for evaluation of police behavior (e.g., stating the reason for the stop) across massive datasets.

### Scaling and Future Work

- The creation of a normalized data infrastructure was necessary to handle varied departmental metadata formats
- The tools aim to allow departments to proactively evaluate training and foster better community relationships.

![Screenshot at 00:04: Introduction slide listing the main PIs and Co-PIs for the project.](https://ss.rapidrecap.app/screens/fZC2eX-RGts/00-00-04.png)
![Screenshot at 01:36: A slide showing a split-screen transcript of the initial interaction between Officer Lane and George Floyd, highlighting the dialogue.](https://ss.rapidrecap.app/screens/fZC2eX-RGts/00-01-36.png)
![Screenshot at 02:28: A slide categorizing George Floyd's dialogue into emotional/interactive themes: Exclamation, Apology, Plea, Proclaim Innocence, Request for reason, and Explanation.](https://ss.rapidrecap.app/screens/fZC2eX-RGts/00-02-28.png)
![Screenshot at 04:44: An illustration depicting a person behind bars whose form is composed of binary code, symbolizing the impact of surveillance technology on individuals.](https://ss.rapidrecap.app/screens/fZC2eX-RGts/00-04-44.png)
![Screenshot at 08:45: A graph illustrating that police officers use more respectful language \(higher 'Respect' score on the y-axis\) with White drivers than Black drivers over the duration of the interaction \(Time in Interaction on the x-axis\).](https://ss.rapidrecap.app/screens/fZC2eX-RGts/00-08-45.png)
![Screenshot at 09:03: A list detailing things officers said more to White drivers, such as 'All right, sir, take care' and 'I just want you and your baby to be safe.'](https://ss.rapidrecap.app/screens/fZC2eX-RGts/00-09-03.png)
![Screenshot at 09:46: A slide summarizing the first 45 words of officers' speech, showing they are less likely to state the reason for the stop and more likely to issue an order to Black drivers.](https://ss.rapidrecap.app/screens/fZC2eX-RGts/00-09-46.png)
![Screenshot at 11:01: A graphic illustrating the scale of data used: 1,000 traffic stops leading to 1.3 million pieces of footage \(550 terabytes of data\).](https://ss.rapidrecap.app/screens/fZC2eX-RGts/00-11-01.png)
![Screenshot at 13:16: A diagram showing the data pipeline: raw audio from two departments merges into a Canonical Form, which is then used for Transcription.](https://ss.rapidrecap.app/screens/fZC2eX-RGts/00-13-16.png)
![Screenshot at 21:22: A bar chart showing that Black Americans in redlined areas \(Area 1 & 5\) report significantly worse treatment \(34.94%\) compared to Black Americans in non-redlined areas \(43.29%\).](https://ss.rapidrecap.app/screens/fZC2eX-RGts/00-21-22.png)
![Screenshot at 33:29: A slide summarizing the challenges in scaling the project, requiring professional, background-checked transcribers and highly-trained expert labelers.](https://ss.rapidrecap.app/screens/fZC2eX-RGts/00-33-29.png)