AI Body Camera Analytics Had No Effect on Troopers, Trial Finds

Artificial intelligence that reviews police body-worn camera footage did not change how troopers in a large state police agency behaved in the field, according to a randomized controlled trial conducted with the Arizona Department of Public Safety and published in the journal Crime Science. The six-month experiment found no statistically significant effects on trooper stops, citations, arrests, use of force, or complaints after correcting for multiple comparisons, and only a modest, statistically nonsignificant increase in a machine-generated measure of professionalism among troopers who had access to the platform. The authors conclude that the technology produced neither the hoped-for “civilizing effect” on officer behavior nor the feared “de-policing” reduction in activity.

Once recordings are uploaded to an agency’s storage system, the algorithm creates a transcript of each conversation, classifies incidents based on what was said, and surfaces footage according to pre-set criteria. Potentially negative behaviors, such as impolite language, directed profanity, or use of force, are flagged for supervisor review, while positive behaviors, such as attempts to de-escalate a community member or displays of high composure, are also identified. The platform assigns a professionalism rating of High, Standard, or Substandard to every recorded encounter.

The motivation for the research lies in what policing researchers call the BWC footage review problem. Agencies deploying body cameras quickly discovered that their officers record far more footage than anyone can watch. Although some footage is subject to automatic review, such as recordings involving use of force, or to random audits, the vast majority, 95 percent or more, is never reviewed by anyone. In one illustrative case, only 3.8 percent of nearly 16,000 videos uploaded by 300 Glendale, Arizona officers in a single month were reviewed internally or by someone outside the department. The failure to review footage undermines the core promised benefits of body cameras, because the civilizing effect depends on officers believing their behavior will actually be observed.

AI-driven analytics has emerged as a possible solution, promising to subject 100 percent of uploaded footage to automated review, identify both problematic and exceptional behavior, support supervisory evaluation, and even facilitate report writing. Interest among law enforcement agencies has grown rapidly, and Truleo reports that its product is trusted by more than 1,000 departments. Yet the evidence base remains thin, and the technology has drawn opposition from police unions, some of which have described the analytics as spying on officers, and from civil rights groups including the ACLU of Colorado. Concerns center on surveillance: constant workplace monitoring has been shown in general research to produce declining employee motivation and reduced productivity, effects that in a policing context would amount to de-policing, a reduction in officer activity under intense scrutiny.

To test these competing possibilities, the Arizona Department of Public Safety deployed Truleo in a two-phase randomized controlled trial. A pilot phase involving 37 troopers in two Phoenix-area districts ran from December 9, 2024 to June 9, 2025, followed by a main phase with 277 troopers across the remaining 18 districts statewide from April 14, 2025 to October 14, 2025. In total, the researchers analyzed 314 troopers, with 155 assigned to the Treatment group and 159 to the Control group, stratifying the randomization by sex. Troopers in the Treatment group and their sergeants received real-time continuous access to Truleo data, while Control group troopers and their sergeants had no access, although Truleo data was still generated for them in the background so the researchers could evaluate it. Strict department-level permissions were set with Truleo staff to prevent contamination.

The outcomes included six-month counts of self-initiated stops, citations, arrests, use of force, and complaints for each trooper, drawn from administrative data aggregated to the officer level. The researchers used negative binomial regression models with robust standard errors and applied a permutation-based stepdown method to control the family-wise error rate, reporting both adjusted and unadjusted p values. The analysis plan was pre-registered on the Open Science Foundation. Treatment group troopers averaged 421.23 stops during the study period compared with 409.47 for Control troopers, along with 289.91 citations versus 267.15, 17.79 arrests versus 16.75, 0.56 uses of force versus 0.39, and 0.12 complaints versus 0.06. None of these differences reached statistical significance after multiple-comparison correction. There was a nominally significant increase in complaints among Treatment troopers in the main trial, with an incident rate ratio of 3.318 and an unadjusted p value of .028, but this finding did not survive correction, and complaints were a rare outcome.

On the professionalism side, Truleo assigned ratings to 164,383 encounters across both phases. The High Professionalism label is applied when an officer uses 25 or more words of explanation before an official action is taken, there is no threat of force, no force is used, and no below-standard language appears. Substandard ratings include directed profanity or insults toward a community member. Only three encounters in the entire study were rated Substandard, so the researchers focused on High versus Standard professionalism using mixed effects linear regression models that account for the nesting of videos within officers. In the main trial, 14 percent of Treatment group encounters received a High Professionalism label compared with 13 percent for the Control group, and the pooled sample showed 14 percent versus 12 percent. The regression models showed consistent nonsignificant coefficients of approximately one percentage point favoring the Treatment group. The authors note that in the pooled analysis this difference translates to an approximate relative increase of 8.39 percent in High Professionalism encounters, or roughly 1,746 more such encounters, a difference they suggest may still be practically meaningful even though it does not provide strong support for a civilizing effect.

The study also examined whether Truleo’s machine-generated professionalism measure aligns with traditional performance metrics, an important question because the measure has not been independently validated. The correlation analysis found statistically significant but weak positive associations between High Professionalism and stops (r = 0.16, p < 0.01) and citations (r = 0.13, p < 0.05), and near-zero relationships with arrests, use of force, and complaints (r < .10, p > 0.05). The researchers caution that directionality cannot be parsed out: either skilled communicators make more stops and issue more citations, or troopers who make more stops and citations learn better communication skills. The authors also identify what they call a fundamental flaw in the High Professionalism rating, which excludes any use of force by definition despite force being a core and sometimes necessary feature of the police role. They further raise the possibility of intentional gamification, noting that officers in prior studies were acutely aware of the 25-word criterion, and asking whether officers simply add a few relatively meaningless statements in a form of metric hacking to secure a High rating. A recent study found that independent observers could identify Substandard professionalism but had difficulty distinguishing High from Standard ratings.

The authors offer several explanations for the null results. The study was underpowered to detect small and moderate effects at the officer level for the activity measures, with minimum detectable effects ranging from an incident rate ratio of 1.19 for stops to 2.67 for complaints. It is also possible that problematic behavior is simply rare in the Arizona Department of Public Safety, a professional and innovative agency whose director is the immediate past president of the National Organization of Black Law Enforcement Executives and which volunteered for the experiment. The agency’s unique mission may matter as well: all study troopers perform traffic duties as a central responsibility, so any Truleo effect concentrated among traffic officers might be washed out across the whole force. In the authors’ earlier evaluation in two mid-sized Arizona municipal departments, Apache Junction and Casa Grande, only two of ten activity outcomes reached significance and both became nonsignificant after multiple-comparison adjustment. The researchers also acknowledge possible informal contamination, since Control troopers knew Truleo was active, worked in shared squads with Treatment colleagues, and knew they would eventually gain access themselves, which could have attenuated the treatment effect through spillover or what has been called contagious accountability.

The findings arrive as adoption of AI-driven body camera analytics accelerates despite a limited evidence base. The authors emphasize that the benefits of such platforms likely extend beyond officer field behavior, including dramatically increasing footage review, enhancing supervision, identifying negative and exceptional conduct, and facilitating report writing, and they suggest agencies should investigate those benefits when considering adoption. They call for future research to formally validate vendor-generated measures, to include traditional activity measures as outcomes while recognizing their limitations, to explore community-member outcomes and situational factors that the current study could not capture, and to develop more nuanced measures of the skills that define good policing.

Subject of Research: A randomized controlled trial testing the effect of AI-based body-worn camera analytics on state trooper behavior and performance measures

Article Title: Does AI Generate a Civilizing or De-policing Effect? Testing the Impact of AI-Based Body Worn Camera Analytics in a Large State Police Agency

Article References: White, M. D., Watts, S., & Malm, A. (2026). Does AI Generate a Civilizing or De-policing Effect? Testing the Impact of AI-Based Body Worn Camera Analytics in a Large State Police Agency. Crime Science, 15(1), Article 36. https://doi.org/10.1186/s40163-026-00293-x

Image Credits: AI Generated

DOI: 10.1186/s40163-026-00293-x

Keywords: body-worn cameras, artificial intelligence, policing, randomized controlled trial, Truleo, de-policing, civilizing effect, law enforcement accountability, natural language processing, officer professionalism, Arizona Department of Public Safety, police surveillance