Auto-Drafting Police Reports from Noisy ASR Outputs: A Trust-Centered LLM Approach

Param Kulkarni, Yingchi Liu, Hao-Ming Fu, Shaohua Yang, Isuru Gunasekara, Matt Peloquin, Noah Spitzer-Williams, Xiaotian Zhou, Xiaozhong Liu, Zhengping Ji, Yasser Ibrahim

Abstract

Achieving a delicate balance between fostering trust in law enforcement and protecting the rights of both officers and civilians continues to emerge as a pressing research and product challenge in the world today. In the pursuit of fairness and transparency, this study presents an innovative AI-driven system designed to generate police report drafts from complex, noisy, and multi-role dialogue data. Our approach intelligently extracts key elements of law enforcement interactions and includes them in the draft, producing structured narratives that are not only high in quality but also reinforce accountability and procedural clarity. This framework holds the potential to transform the reporting process, ensuring greater oversight, consistency, and fairness in future policing practices.

Introduction

Police officers currently dedicate a significant portion of their shifts—often up to 40%—to administrative tasks, primarily report writing. This burden contributes to burnout and reduces the time available for community engagement and active policing. While Large Language Models (LLMs) offer a potential solution for automating documentation, the law enforcement domain presents unique challenges. Reports must be generated from noisy body-worn camera (BWC) audio, which includes environmental interference and overlapping dialogue. Furthermore, the "trust gap" necessitates a system that avoids "hallucinations" and prevents automation bias, ensuring that legal conclusions remain the sole responsibility of the human officer.

System Architecture and Methods

The proposed system employs a multi-stage, human-in-the-loop (HITL) architecture designed to maximize accuracy while maintaining strict accountability.

Pre-Generation Metadata

To prevent the AI from making conclusory legal statements, the system requires officers to manually input the incident type and severity (e.g., "Public Disorder" and "Misdemeanor") before any narrative is generated. This ensures the LLM acts as a scribe rather than a judicial arbiter.

Denoising and Synthesis

The core engine utilizes state-of-the-art LLMs to process noisy Automatic Speech Recognition (ASR) transcripts. These models are prompted to extract essential elements and synthesize them into a coherent narrative, effectively "denoising" the messy dialogue often found in BWC recordings.

Intentional Friction: The [INSERT] Feature

A critical design choice to combat automation bias is the inclusion of "intentional friction." When the system identifies gaps in the transcript or areas requiring specific officer observation, it places [INSERT] tags within the draft. The workflow prevents submission until the officer has manually addressed these specific points.

The end-to-end workflow from body-worn camera recording to final report signing.

Final Verification

The process concludes with a "Review and Sign" stage. The officer, as the "author of record," must review and edit the draft before providing a digital signature.

Extracted figure from page 3 of Auto-Drafting Police Reports from Noisy ASR Outputs: A Trust-Centered LLM Approach

Evaluation and Results

The system was evaluated using pilot data from 326 agencies and a double-blind expert study involving 113 report pairs.

Efficiency Gains

Officers reported a significant reduction in administrative burden, saving an average of 21.93 minutes per report—a 41.8% decrease in writing time. User satisfaction was high, with an average rating of 4.46/5.

Quality and Objectivity

In double-blind evaluations by 24 law enforcement equity experts, AI-assisted reports statistically outperformed manual reports in terminology (4.20 vs. 3.97, p=0.033) and coherence (4.05 vs. 3.75, p=0.019). Crucially, the system maintained parity with human-written reports in terms of neutrality and objectivity, indicating that safety-focused prompting successfully mitigated potential AI bias.

Discussion and Conclusion

This research demonstrates that LLMs can be effectively deployed in high-stakes environments by prioritizing a trust-centered design. By treating the AI as a tool for synthesis rather than a decision-maker, the system improves report quality and officer efficiency without sacrificing procedural justice.

Future work will address ongoing ASR bottlenecks in extreme noise environments and investigate the long-term impact of AI-assisted documentation on court proceedings.

A demonstration video of this system is available here.

Summary updated

Abstract

Achieving a delicate balance between fostering trust in law enforcement and protecting the rights of both officers and civilians continues to emerge as a pressing research and product challenge in the world today. In the pursuit of fairness and transparency, this study presents an innovative AI-driven system designed to generate police report drafts from complex, noisy, and multi-role dialogue data. Our approach intelligently extracts key elements of law enforcement interactions and includes them in the draft, producing structured narratives that are not only high in quality but also reinforce accountability and procedural clarity. This framework holds the potential to transform the reporting process, ensuring greater oversight, consistency, and fairness in future policing practices. A demonstration video of our system can be accessed at https://drive.google.com/file/d/1kBrsGGR8e3B5xPSblrchRGj-Y-kpCHNO/view?usp=sharing
BibTeX
@inproceedings{kulkarni2025auto,
  author = {Param Kulkarni and Yingchi Liu and Hao-Ming Fu and Shaohua Yang and Isuru Gunasekara and Matt Peloquin and Noah Spitzer-Williams and Xiaotian Zhou and Xiaozhong Liu and Zhengping Ji and Yasser Ibrahim},
  title = {{Auto-Drafting Police Reports from Noisy ASR Outputs: A Trust-Centered LLM Approach}},
  booktitle = {The Web Conference},
  year = {2025},
  url = {https://openreview.net/forum?id=h4GhrSLeEY},
}
Figure