Gunshot detections systems work. No, gunshot detection systems don’t work. Can we have it both ways? It’s kind of like the first line in A Tale of Two Cities: “It was the best of times, it was the worst of times.” Well? Which is it? This week we’ll be discussing Gunshot Detection Systems and the impact on evidence collection and crime. Dr. Eric Piza, the Lipman Family Professor of Criminology & Criminal Justice and Director of Crime Analysis Initiatives at Northeastern University, joins the podcast to discuss his 2024 study, “Gunshot detection technology effect on gun violence in Kansas City, Missouri: A microsynthetic control evaluation, coauthored by David N. Hatten, George O. Mohler, Jeremy G. Carter, and Jisoo Cho.
Exploring Officer Patrol Behaviors Using Automated Vehicle Locator and Body-Worn Camera Data
Funder: National Institute of Justice ($862,804)
Principal Investigator: Eric L. Piza
Co-Principal Investigators: Nathan T. Connealy & Victoria A. Sytsma
Project Overview
This project is a mixed-methods research project that leverages automated vehicle locator (AVL) data and body-worn camera (BWC) video to analyze police patrol activity and police-citizen interactions. We will first operationalize daily patrol officer travel patterns from AVL data, specifically focusing on committed vs. uncommitted time, proactive vs. reactive deployment, driving speed, single vs. multiple-vehicle responses, and stationary vs. in-motion presence. We will then select a stratified random sample of incidents for further analysis through systematic social observation (SSO) of associated BWC video. The SSO will measure the degree to which police officers adhere to principles of procedural justice, deploy de-escalation tactics in contentious citizen encounters, and use their discretion to minimize punitive responses. Data from the AVL and BWC video will further be triangulated with a wide range of KCPD and public data sources to identify contextual factors that are significantly related to a range of police patrol behaviors and response outcomes. This study’s novel methodology provides a blueprint for leveraging AVL and BWC data in support of evidence-based policing and police reform.
Project Publications
AI vs. Human Transcription: Evaluating Accuracy and Meaning in Police Body-Worn Camera Footage
Savannah A. Reid, Stephen Abeyta, Eric L. Piza, Nathan T. Connealy, and Victoria A. Sytsma
Journal of Experimental Criminology (2026)
Key Takeaways
- AI-generated transcripts of body-worn camera footage capture most of the same words as human-edited transcripts, but are much worse at identifying who is speaking
- AI transcripts identify only about two-thirds as many speakers as human-edited transcripts on average (3.3 vs. 4.8 per document), and never detect more than 7 speakers even when humans identify as many as 16
- AI tends to consolidate or drop less-prominent voices (bystanders, radio traffic, secondary civilians), collapsing back-and-forth exchanges into fewer, longer speaker turns
- Document-level content similarity (the overall words captured) stays high even when speaker attribution is poor; AI can “get the words right” while “getting people wrong”
Research Summary
Body-worn cameras (BWCs) are now standard in American policing, but the resulting volume of footage makes systematic review difficult without automation. AI-powered automatic speech recognition (ASR) promises to convert this footage into searchable, analyzable text at scale, and vendors such as Axon have already partnered with ASR providers (Rev.ai) to offer automated transcription to law enforcement clients. But the value of any downstream AI-driven analysis of BWC footage depends entirely on whether the underlying transcript accurately captures what was said and who said it.
This study uses a sample of 73 police incidents (176 BWC videos, roughly 23 hours of footage) from the Kansas City, Missouri Police Department. Each video’s audio was transcribed twice through Rev.ai: once through fully automated AI transcription, and once through Rev’s human-edited service, in which professional transcribers review and correct the AI draft.
The analysis proceeded in three steps: extracting each speaker and their quotations from both transcript types, comparing how closely individual speaker profiles aligned across the AI and human-edited pairs (using Jaccard and cosine text-similarity metrics, including bi-gram/tri-gram sequencing), and comparing whole-document similarity to assess overall lexical overlap.
The clearest finding is a systematic under-counting of speakers by the AI system. AI-generated transcripts identified only about two-thirds as many unique speakers as human-edited versions (an average of 3.3 versus 4.8 per document), never surpassing 7 speakers even in encounters where human transcribers identified as many as 16. AI transcripts tended to consolidate the conversation into fewer speakers with longer, merged quotations. Despite this, whole-document similarity scores between AI and human-edited transcripts were comparatively high.
Together, these patterns indicate that AI transcription accurately captures content but struggles with attribution, especially in complex, multi-party scenes with background noise, overlapping speech, or less-prominent voices such as bystanders and radio traffic. This distinction matters because different uses of BWC transcripts have very different tolerance for attribution error. Content-level applications can likely rely on AI transcription alone, since aggregate word overlap is high and individual misattributions are unlikely to skew large-sample results. But applications that hinge on knowing who spoke, how often, and in what sequence—such as internal investigations, use-of-force reviews, officer performance evaluations, and procedural-justice or de-escalation research—require accurate speaker attribution and are vulnerable to being compromised by AI-only transcripts.
These results extend a broader body of ASR research showing that speech-recognition accuracy varies substantially across audio environments and speaker characteristics. The study also speaks to rapidly developing criminal justice research using AI to score BWC footage for officer professionalism and communication quality, which has generally found that AI-derived behavioral codes can correlate with human ratings and even reproduce treatment effects from earlier trials—but with notable discrepancies on finer-grained distinctions and across software platforms. Where most of that literature evaluates AI performance at the level of behavioral codes or outcomes, this study shifts the focus to the transcription layer those codes ultimately depend on. In doing so, the study adds an important qualification to an otherwise optimistic literature on AI-assisted BWC review: alignment at the outcome level does not guarantee that the underlying text faithfully represents who spoke and how an interaction actually unfolded.
Computational Approaches and the Future of Urban Crime Research
Gian Maria Campedelli, Zubin Jelveh, Aaron Chalfin, Daniel Semenza, Eric Piza, Ariadna Albors Zumel, Bruno Lepri, and Patrick Sharkey
Nature (2026)
Abstract
Urban environments have long been a central focus for crime researchers across diverse disciplines. Over the past few decades, this heterogeneous area of inquiry has experienced substantial methodological and empirical change, driven by the emergence of novel datasets and the increasing use of flexible computational methods. In this Review, we take stock of this evolution and examine the potential that these developments hold for advancing urban crime research, while also addressing the persistent challenges that continue to shape the field. Building on this overview, we emphasize the promise that computational methods offer for more rigorous causal inference beyond traditional prediction tasks. Finally, we outline three key directions for future research to ensure that new data and computational tools are used effectively: greater integration across disciplines, improved open science standards and a broader scope of inquiry beyond Western contexts. In doing so, we aim to support more rigorous research and inform the development of more effective policies for safer and more sustainable cities worldwide.
Geographic Networks of Gun Violence: Exploring Hot Spot Connectivity through Ballistic Evidence and Social Network Analysis
Eric L. Piza, Cassie McMillain, and Daniel Trovato
Journal of Research in Crime & Delinquency (2026)
Key Takeaways
- Ballistic evidence (NIBIN hits) reveals spatial links between gun violence hot spots
- A single large gun violence network spans much of the city, with 53% of street segments with NIBIN evidence connected to each other
- The gun violence network reflects patterns of hot spot homophily, with gun violence hot spots more likely to connect to other hot spots
- Hot spot homophily remains statistically significant even after controlling for geographic proximity, neighborhood demographics, built environment features, and police enforcement activity
- Because gun violence hot spots are interconnected, intervening in highly connected hot spots can generate residual deterrence effects that ripple through the broader network
Research Summary
The criminology of place has long established that crime is not evenly distributed across a city. A small minority of geographic hot spots account for a disproportionate share of criminal activity. Despite this well-documented concentration, most crime-and-place research implicitly treats hot spots as independent, self-contained units. Environmental criminology, including crime pattern theory, has long emphasized that the interconnectedness of places shapes how crime opportunities are presented to motivated offenders. A social network lens offers a promising way to formalize and test these interconnections empirically.
This study analyzes patterns of gun violence in Kansas City, Missouri, leveraging data from the National Integrated Ballistic Information Network (NIBIN). NIBIN collects ballistic imaging data from spent projectiles and cartridges recovered at crime scenes or test-fired from seized firearms. When ballistic evidence from the same firearm is recovered at two different locations, NIBIN generates a “hit” — a confirmed link between those two places.
NIBIN hits are used to construct a geographic network of gun violence in Kansas City. Our analytic sample includes 1,701 street segments connected by 2,248 edges derived from 1,349 unique crime guns with at least one NIBIN hit from 2014 to 2019. Street segments are then classified as hot spots, low gun crime, or no gun crime places.
Exponential random graph models (ERGMs) test whether hot spots are disproportionately connected to one another. This method compares the network’s structural patterns against thousands of randomly simulated networks to determine whether observed patterns exceed random chance. Models include a range of control variables covering geographic proximity, the built environment, neighborhood demographics, police enforcement, and endogenous network processes.
Over half of all street segments with NIBIN evidence belong to a single large connected component. This indicates that gun violence in Kansas City is not fragmented across isolated pockets but is structurally linked across a substantial portion of the city’s geography. Visual inspection of the network confirms that hot spots tend to cluster in the central core of this connected component, while lower-crime segments occupy the periphery.
Gun violence hot spots exhibit strong homophily — that is, they are significantly more likely to be connected to other hot spots than would be expected at random. When two street segments are both hot spots, the odds that ballistic evidence from the same crime gun appeared in both locations increases by approximately 43–47%. This pattern, which we term hot spot homophily, holds consistently across all model specifications. A separate set of mixing models further confirms that connections between pairs of hot spots are significantly more likely to occur in the observed network than any other pairing — hot spot to low gun crime street segment, hot spot to no gun crime crime street segment, or two low crime street segments.
This study advances the criminology of place by demonstrating that gun violence hot spots are not self-contained but are embedded within an interconnected geographic network. These findings have meaningful implications for place-based policing. If hot spots are interconnected, then successfully intervening in a highly connected hot spot may produce spillover prevention effects throughout the broader gun violence network. Law enforcement can leverage NIBIN data to prioritize intervention sites, focusing on places that serve as central hubs in the gun violence network.
Officers’ body cameras all fell off during violent arrest. Phoenix doesn’t know how much it happens
Despite tens of millions of dollars spent on the technology, the Phoenix Police Department does not track how often officers’ body cameras fall off during use-of-force incidents.
ABC15 asked the city if it kept data on “body-worn camera dislodgements” while reporting on a violent and controversial arrest that appears to be set to end up as a federal lawsuit.
A Phoenix police spokesperson said it does not and does not believe it’s a concern.
Eric L. Piza Appointed Lipman Family Professor of Criminal Justice
On February 2nd, Northeastern University celebrated the installation of Eric L. Piza as the Lipman Family Professor of Criminal Justice. The endowed professorship recognizes Professor Piza’s leading scholarship in evidence-based policing, place-based crime prevention, and the use of technology and data in criminal justice—and strengthens the School of Criminology and Criminal Justice’s commitment to research that informs policy and practice.
The Lipman Family Professor of Criminal Justice is an endowed position made possible by the Lipman family’s support for Northeastern’s School of Criminology and Criminal Justice. When the fund was established, it specified that the occupant of the chair be an outstanding and distinguished scholar in criminology and criminal justice. Ben Lipman, a graduate of the College of Criminal Justice and a former member of the Northeastern University Corporation, represented the Lipman family at the installation event.
KC wants to keep funding gunfire detection system. But has it reduced crime?
The Kansas City Police Department continues to request funding for a gunshot detection technology that research shows has not reduced shootings or improved case clearance rates. While the technology has not improved clearance rates or reduced shootings, research has shown that it has improved response times and evidence recovery.
Here Be Dragons: Burdens of Knowledge and Innovation in Evidence-Based Policing
Eric L. Piza
Evidence Base: Criminal Justice Research, Policy and Action (2026)
Key Takeaways
- Evidence-based policing must evolve beyond “what works” to provide more value to police practitioners
- Innovation is constrained by the growing “burden of knowledge” that typifies developed sciences
- Further innovation requires deeper and more resource-intensive research to generate meaningful advances
- The real-world impact of evidence-based strategies depends more on implementation quality, organizational capacity, and local context than on program design alone.
- Sustaining evidence-based policing requires investment in a broader knowledge infrastructure that integrates evaluation, implementation science, and officer-level data into routine decision-making
Research Summary
This essay argues that evidence-based policing (EBP) has reached a critical inflection point. Decades of rigorous research have clarified which policing strategies can reduce crime, but this success has also produced new challenges. The accumulation of knowledge has created a “burden of knowledge,” making further innovation harder and leaving many aspects of policing practice poorly understood. I contend that the next phase of EBP must move beyond asking what works and focus instead on how, why, and under what conditions policing strategies succeed or fail.
Research has consistently shown that proactive strategies outperform reactive ones, that focusing resources on high-risk places and people is more effective than spreading them broadly, and that problem-oriented approaches outperform generic enforcement. Systematic reviews now provide strong evidence supporting strategies such as hot spots policing, problem-oriented policing, and focused deterrence. As a result, policing is no longer a low-information environment.
Ironically, this growth in evidence has widened the gap between research and practice. Police agencies often struggle to implement evidence-based strategies effectively, even when strong evidence exists. The “burden of knowledge” mechanism explains that as a field matures, advancing it requires greater effort and more complex forms of inquiry. In policing, what police leaders increasingly need is guidance on implementation, adaptation, organizational capacity, and local context.
The essay does not argue for abandoning rigorous impact evaluations. Rigorous designs remain essential for determining effectiveness. However, an exclusive focus on causal outcomes risks stifling innovation by privileging a narrow set of research questions and methods. Further advancement requires a second generation of evidence-based policing built on a broader knowledge infrastructure.
A second-generation EBP should have three priorities.
First, implementation science is essential for understanding how evidence-based practices are adopted, adapted, and sustained in real-world agencies. Factors such as leadership, organizational culture, resources, officer motivation, and external pressures strongly shape outcomes and must be studied systematically.
Second, EBP must better track officer activity. Without knowing what officers actually do on the street, agencies cannot determine whether outcomes reflect strategy design or execution. Technologies such as body-worn cameras and automated vehicle locators provide unprecedented opportunities to study officer behavior, treatment dosage, and police–community interactions.
Third, scholars should better embrace basic and descriptive research. Exploratory, diagnostic, and qualitative studies often generate the foundational knowledge that enables innovation, even if they rank lower on traditional methodological hierarchies.
In conclusion, evidence-based policing stands at a pivotal moment. The easy questions about what works have largely been answered, but the harder work of understanding how policing functions in practice remains. Addressing this challenge requires embracing methodological diversity and treating implementation, context, and officer behavior as central to evidence-based policing’s future.
Pokémon cards bring business — and thieves
When Ron Zeida woke up in the middle of the night to a barrage of texts, he knew something was wrong.
On Dec. 1, a burglar broke into his Vanguard Comics store in Barnstable and swiped roughly $1,000 worth of merchandise from the shelves. In under a minute, the thief left the store’s glass door shattered and its tight-knit community shaken. Three weeks later, police are still combing for leads, Zeida said.
But given the range of merchandise that Vanguard carries, it could have been much worse.
Eric L. Piza, Ph.D. 