The Subjectivity of Respect in Police Traffic Stops: Modeling Community Perspectives in Body-Worn Camera Footage

TL;DR

This work tackles the subjectivity of respect in police–civilian traffic stops by constructing the LAPD Respect Dataset, a large-scale corpus of BWC transcripts annotated for officer and civilian respect from multiple community perspectives. It develops a domain-specific rubric (Emotions, Professionalism, Communication, and Contextual Moderators) and a rubric-aware LLM framework to evaluate rationales and produce rubric-grounded preference data for perspective alignment. By combining supervised fine-tuning with Direct Preference Optimization (DPO) and perspective conditioning, the approach yields improved rating accuracy and rationale quality across annotator groups, with the strongest gains for justice-system-impacted participants. The results demonstrate that modeling perspective-aware respect improves interpretability and trust in high-stakes policing contexts and offers a path toward more inclusive AI systems that reflect diverse community expectations.

Abstract

Traffic stops are among the most frequent police-civilian interactions, and body-worn cameras (BWCs) provide a unique record of how these encounters unfold. Respect is a central dimension of these interactions, shaping public trust and perceived legitimacy, yet its interpretation is inherently subjective and shaped by lived experience, rendering community-specific perspectives a critical consideration. Leveraging unprecedented access to Los Angeles Police Department BWC footage, we introduce the first large-scale traffic-stop dataset annotated with respect ratings and free-text rationales from multiple perspectives. By sampling annotators from police-affiliated, justice-system-impacted, and non-affiliated Los Angeles residents, we enable the systematic study of perceptual differences across diverse communities. To this end, we (i) develop a domain-specific evaluation rubric grounded in procedural justice theory, LAPD training materials, and extensive fieldwork; (ii) introduce a rubric-driven preference data construction framework for perspective-consistent alignment; and (iii) propose a perspective-aware modeling framework that predicts personalized respect ratings and generates annotator-specific rationales for both officers and civilian drivers from traffic-stop transcripts. Across all three annotator groups, our approach improves both rating prediction performance and rationale alignment. Our perspective-aware framework enables law enforcement to better understand diverse community expectations, providing a vital tool for building public trust and procedural legitimacy.

Figures (8)

• Figure 1: Overview of the proposed framework, illustrating multi-perspective respect annotation from BWC transcripts, and domain-specific rubric development. The framework integrates rubric-guided preference data synthesis with supervised fine-tuning and alignment to produce perspective-aware respect ratings and rationales.
• Figure 2: Rating MAE (lower is better) and $F_1(\hat{\rho})$ (higher is better) for officers (top) and drivers (bottom).
• Figure 3: Our annotation platform developed for annotating BWV data.
• Figure 4: Rubric and instructions used by the LLM-as-a-judge to evaluate officer-respect rationales and output structured rubric activations.
• Figure 5: Rubric and instructions used by the LLM-as-a-judge to evaluate driver-respect rationales and output structured rubric activations.