John G. Peters, Jr., Ph.D., and John Black, D.B.A.
©2026. A.R.R.
Artificial Intelligence (AI) is reshaping public safety work and nowhere is that shift more visible than in AI-assisted Body-Worn Camera (BWC) report writing.
The term “artificial intelligence” was introduced by John McCarthy and colleagues in a 1955 proposal for the 1956 Dartmouth Summer Research Project on Artificial Intelligence. AI still has no universally accepted definition; Ian McCulloh, Ph.D., of Johns Hopkins University, offers a working one: “a machine that can sense, reason and act” (personal communication). What began as an academic research field seven decades ago is now a rapidly evolving technology which agencies and researchers are working to understand and evaluate.
Law enforcement agencies are exploring AI across a wide range of technologies and tasks, including AI-enabled drones and robotic systems, analytics, transcription, and report writing. In report writing and investigative settings, AI-assisted tools can generate transcripts, organize recorded information, flag possible inconsistencies for human review, and help users reconstruct events from recorded information.
Human in the Loop
That growing reliance comes with an important requirement: human review. AI generated content can contain errors, omissions or misleading information, and someone must remain responsible for verifying its accuracy. In AI-assisted BWC reporting, the officer remains responsible for reviewing and approving the report, while supervisors may conduct additional review as part of their normal responsibilities. That review can add to existing workloads and, where job duties materially change, may also raise labor management issues requiring discussion or bargaining.
A Need for Scientific Study
Despite this rapid adoption and the operational and labor questions it raises, little scientific research has examined how well AI-assisted BWC reporting actually works and its use in the workplace. Does it save employees time? What concerns do they have about using it? Do agencies have written policies governing their use? And what training, if any, is given before employees rely on it?
To begin answering these questions, the Institute for the Prevention of In-custody Deaths, Inc. (IPICD), through its Center of Excellence for Artificial Intelligence (AI Center), conducted a national survey distributed independently by E-mail through the California Force Instructors Association, Police1, Xiphos, and other professional interest channels.
The 18 question survey was reviewed for content relevance prior to being E-mailed to individual law enforcement personnel across the United States. Select descriptive statistical findings are presented in this article. Space constraints prohibit listing and discussing each finding. A complete finding report with both quantitative (descriptive and inferential statistics) and qualitative analyses is forthcoming and will be published by the IPICD.
Survey Methodology
The survey was developed using SurveyMonkey and peer reviewed before being distributed through law enforcement professional interest channels with an estimated combined audience of approximately 225,000 individuals. A total of 214 usable survey responses were received from 34 states, the District of Columbia and Puerto Rico. Not every respondent answered every question, so the number of responses (n) varies by survey item. All responses were anonymous and are reported in aggregate. The survey was conducted for research and educational purposes.
Using the estimated distribution audience of 225,000 as the denominator, the 214 usable responses represent approximately 0.1 percent of that audience. Because the survey was distributed through multiple professional interest channels and respondents opted in rather than being randomly selected, the findings describe this self-selected group and should not be interpreted as representative of American law enforcement as a whole.
Respondent Characteristics
The largest single age group, 43 to 54, comprised 48 percent of respondents; another 33 percent were 55 and older, and only 1.9 percent were 25 to 30 years old (n=214). Supervisors – including sergeants, lieutenants, captains, and higher ranking personnel – accounted for 49 percent of respondents, while patrol officers and deputies accounted for 13 percent (n=212). The specific subgroup size (n) varies throughout the survey because not every respondent answered every question.
AI-assisted Report Writing Platforms
Among the 73 respondents who answered the question about AI-assisted report writing platforms, 40 percent reported using Axon Draft One and 38 percent reported using another AI-assisted report writing platform. In a separate question about agency authorization, 34 percent said their agency authorized the use of an AI-assisted report writing tool, 57 percent said their agency did not and nine percent were unsure (n=191).
Policy and Training
When asked whether their agency had a written policy governing the use of AI to assist with report writing, 62 percent replied “yes,” 31 percent replied “no” and seven percent were unsure (n=96). Among the 78 respondents who answered the training question, 47 percent said their agency had trained them to use AI-assisted report writing tools.
Among respondents for whom both policy and training information were available, agencies reporting having a written AI policy also reported training at a rate of roughly 60 percent, compared with about 15 percent among agencies reporting no policy. This was a strong association within the survey sample, but it does not establish that having a policy causes training; that training leads to policy adoption; or that either is responsible for the other. Other factors, such as agency size, available resources or leadership priorities, may influence both. The forthcoming full report will examine this relationship in greater detail.
Using AI at Work
Among the 77 respondents who answered the question about frequency of use, ten percent said they used an agency authorized AI report writing platform daily; 12 percent used one two to three times per week; 25 percent used one occasionally – a few times per month or less – and 53 percent said they never used one. Open-ended responses also indicated that some respondents used AI tools at work even when those tools were not agency approved, a practice often referred to as “shadow AI.”
In a separate item (n=210), 21 percent of respondents reported using an agency approved AI platform, while 22 percent said they did not use one. Thirty-one percent reported that their agency did not have an approved AI platform and 15 percent reported that their agency did not have a specific policy identifying approved AI platforms.
When usage frequency responses (n=77) were compared with respondents’ reported training status, trained respondents did not report regular use at a clearly higher rate than untrained respondents. The descriptive results suggest that training alone may not explain adoption. Other factors, such as trust in the technology, workload or ease of use, may also play a role, but this survey cannot establish which factors, if any, cause employees to use an AI tool regularly.
How Respondents Use AI at Work
Respondents identified a variety of reasons for using AI at work. Among respondents to this item (n=42), frequently cited benefits included saving time (22 percent); improving report clarity (18 percent); improving spelling, grammar and readability (21 percent); and capturing information the user may have overlooked (13 percent).
AI Use Concerns
Sixty percent of respondents to the report writing concern question (21 of 35) expressed concern about accuracy errors; 51 percent (18 of 35) cited incomplete information; and 46 percent (16 of 35) expressed concern about legal liability if errors went undetected before a report was submitted. Separately, 11 percent of respondents to the time-saving question said AI-assisted report writing did not save them time.
Independent research has also questioned whether AI-assisted report writing necessarily produces the expected time savings. Adams, Barter, McLean, Boehme, and Geary conducted a preregistered randomized controlled trial involving 85 officers and 755 reports in a single police department. The researchers found no statistically significant reduction in report writing time attributable to Axon Draft One in that setting. The finding does not establish that AI-assisted report writing can never save time, but it cautions against assuming that time savings will occur in every agency or application.
Other reporting has raised similar concerns while also illustrating the importance of distinguishing between different AI tools. In July 2026, Forbes reported that officers in Lafayette, Indiana, testing Axon’s Form One – an AI tool designed to help populate administrative report fields – said correcting errors sometimes caused the process to take longer. Form One performs a different function from Draft One’s narrative generation capability, so those experiences should not be treated as a replication of the Draft One study. An ACLU review published in May 2026 cited the Adams study as finding no time savings from Draft One and also cited a separate 2025 evaluation of another AI tool that found increased time per incident.
Legal and prosecutorial responses are also developing. The King County Prosecuting Attorney’s Office in Washington has told area law enforcement agencies that it will not accept police report narratives produced with AI assistance. California requires disclosure when covered AI is used to generate police reports, verification by the report’s author, retention of the original AI generated first draft, and an audit trail. Utah requires an AI disclosure and certification by the author that the AI-assisted report or record was read and reviewed for accuracy.
Technology has also changed in response to concerns about preserving the original AI-generated draft. Axon now provides agencies with the ability to retain original Draft One narratives and records audit events associated with their generation, access and retention. The original draft retention setting is off by default; however, allowing agencies to determine whether and how retained drafts fit their policies and records retention requirements.
When asked more generally about AI use, 72 percent of respondents expressed concern about AI accuracy, while 59 percent were concerned about AI generated reports being submitted without adequate human review (n=212). Respondents also raised concerns involving vendor data security, public trust and employee discipline. An exploratory analysis of these responses suggested two broad patterns: concerns about the accuracy of the work product itself and concerns involving institutional trust and accountability. Because this analysis is preliminary and will be detailed in the Center’s forthcoming full report, the findings should be treated as hypothesis generating rather than conclusive.
Beyond the Percentages: What a Deeper Look Is Beginning to Show
The percentages above tell only part of the story. A deeper statistical review of the same survey responses, now underway as part of the Center’s forthcoming full report, is identifying patterns which warrant additional study. One involves years of service. Respondents with fewer than ten years on the job reported avoiding workplace AI at a rate of approximately 35 percent, or eight of 23 respondents, compared with approximately 13 percent, or 21 of 160 respondents, among those with 17 or more years of service. This descriptive difference should not be generalized beyond the survey sample, but it raises a useful question for future research about whether years of service are associated with willingness to use workplace AI.
A second early finding may be particularly relevant to policy writers. Among respondents for whom both policy and training information were available, agencies reporting a written AI policy also reported training at a rate of roughly 60 percent, compared with about 15 percent among agencies reporting no policy. The relationship does not establish that policy adoption causes training, but it suggests that policy and training frequently occur together. For agency leaders, the practical question is whether implementing an AI policy without accompanying education and training could leave employees uncertain about how the technology should be used.
Leadership and line personnel also reported different perspectives. Supervisors and higher ranking respondents were more likely than line personnel to report knowing their agency’s AI authorization status – approximately 41 percent versus 24 percent (n=111, leadership; n=79, line). Leadership respondents were also more likely to express concern about AI-generated content being submitted without adequate review, approximately 67 percent versus 49 percent among line personnel (n=212). These descriptive differences may reflect different job responsibilities or experiences, but the survey does not establish why the groups responded differently.
Another early result may be relevant to command staff. Among the 77 respondents who answered the usage frequency question, receiving training was not clearly associated with more regular use of an AI tool. Training may expose personnel to the technology, but these survey results do not show that training alone leads to adoption. Other factors – including trust in the tool, workload, perceived usefulness, and ease of use – may also influence whether personnel use AI regularly. Additional analysis will be needed to determine which factors are most closely associated with adoption.
When respondents were grouped according to the concerns they reported about AI – including accuracy, unreviewed content, vendor data security, public trust, and discipline – an exploratory analysis suggested two broad patterns. One centered primarily on institutional trust and accountability, while the other centered primarily on the accuracy of the work product. This preliminary finding will be examined in greater detail in the Center’s forthcoming report and should not yet be interpreted as establishing two definitive categories of law enforcement personnel.
Respondents were asked separately what AI training they wanted for themselves and what training they believed their agencies needed (n=127 and n=119 open-ended responses, respectively). The response patterns differed depending on the question. Responses about personal training tended to emphasize skills and uses extending beyond report writing, while responses about agency needs more often emphasized governance issues such as policy clarity, data security and accountability. Because not every respondent answered both questions, these findings are best understood as differences between the two sets of responses unless a matched response analysis confirms that the comparison involves the same individuals.
A caution belongs here, too: When we broke respondents into smaller groups by rank or job title, some categories were thin, a handful of respondents in certain title categories, for instance, and raw percentages in those small groups should be read as suggestive, not conclusive, until a larger sample confirms them.
There is a broader point worth naming here about method, not just findings. Percentages alone establish that concerns exist and roughly how often. It was only by reading respondents’ own words, not counting checkbox answers, that a distinction emerged: The same people asked for noticeably different things depending on whether the question was about themselves or about their agency. A closed-ended survey question would have missed that shift entirely. Understanding AI adoption in policing, in other words, is not only a technology question; it is also a question about the tacit, on-the-job judgment officers bring to it which shows up in their words long before it shows up in a percentage.
Summary
This survey offers an evidence-based snapshot of how this self-selected group of law enforcement personnel is experiencing AI-assisted report writing, where respondents say policy and training are keeping pace with adoption, where gaps may exist, and what kinds of assistance personnel say they want. It also raises questions for further study: why respondents with fewer years of service reported more AI avoidance than longer-serving respondents; why training status was not clearly associated with regular use in this sample; and whether line personnel and command staff view AI-related risks differently. The AI Center views these findings as starting points for additional research rather than final conclusions.
Readers should therefore treat the findings as descriptive of this survey sample, not as national estimates of law enforcement attitudes or practices. The results identify issues worthy of further discussion and study and provide a foundation for the Center’s forthcoming quantitative and qualitative analyses.
John G. Peters, Jr., Ph.D., serves as President and Chief Learning Officer of the IPICD. He is a graduate of several AI programs from leading institutions including the Massachusetts Institute of Technology Sloan School and Computer Science and Artificial Intelligence Laboratory, University of Las Vegas, and the Whiting School of Engineering, Johns Hopkins University. A former law enforcement administrator and officer, he has been judicially qualified as an expert witness in international, federal, and state courts.
John R. Black, D.B.A., is Director of the IPICD Center of Excellence for Artificial Intelligence. He is a graduate of AI programs, including the Massachusetts Institute of Technology, and his post-doctoral graduate research and studies include metacognition, sensemaking, and decision-making; subjects on which he presents routinely to a range of audiences nationally and internationally. He is a certified thinking leader (CTL) from the Cabrera Research Labs. A retired law enforcement officer (patrol, corrections, investigations, training, and command rank) and a retired US Army Sergeant Major (Special Forces, Psychological Operations, and Civil Affairs), Dr. Black has been judicially qualified as an expert witness in international, federal, and state courts.