Addressing AI in Policing: Recent Concerns and Media Coverage
Addressing AI in Policing: Recent Concerns and Media Coverage
As there have been some recent news stories and comments from academia regarding the use of artificial intelligence to assist in police report document completion, my blog post release this week will discuss and address some of these recent issues. The post will be broken up into two parts, with part two coming out tomorrow.
News Accounts
A news story concerning Axon's Draft One AI report writing tool has recently gained significant attention, though with questionable journalistic research. Initially reported by The Park Record (Park City, Utah) on December 16, 2025, the article described Heber City Police Department's trial of Draft One and a competing product called Code Four. According to the report, the police chief recounted how his son-in-law, an officer at West Jordan PD, encountered a Draft One-generated report claiming the officer had "shape-shifted into a frog" (Taylor, 2025).
The chief is quoted in the article, "'Man, this really looks like an officer wrote it…But when it got to one part it said, And then the officer turned into a frog, and a magic book appeared and began granting wishes....'" (Taylor, 2025). The chief suggested that Harry Potter playing in the background likely affected what was heard and transcribed through the body worn camera (Taylor, 2025).
This incident has since been sensationalized by major outlets including Forbes and Vice News, each presenting conflicting details. Forbes changed key elements, attributing the background noise to "The Princess and the Frog" rather than Harry Potter, and incorrectly stating the report originated from Heber City PD (Daniel, 2026). With contradictory accounts and no official comment from Axon, the true cause remains unclear, a fact largely ignored by AI skeptics using this incident to advance anti-AI positions in policing contexts.
Questions About What Occurred
Without knowing specifically what occurred, let’s look at some of the possibilities. To do that, here’s a simplistic explanation of how Draft One works. An officer on scene has his or her body worn camera activated. When the body camera is deactivated by the officer, the entire audio which was captured is uploaded and then transcribed. After being transcribed, the transcription of the entire audio captured is then passed through the OpenAI large language model (LLM) (with prompting by Axon), to generate a draft report, which the officer is required to review. So, based on this workflow, could movie audio playing in the background while an officer is on scene make its way into the report? Theoretically, yes, but that would represent major issues with the prompt and context engineering for the LLM conducted by Axon.
I tend to believe what occurred is a lack of training and understanding of the Draft One system by these agencies. Within the Axon Draft One User Guide, agencies can activate a setting to “Intentionally Insert Obvious Errors for Removal” (Axon, 2025). The idea behind this is to ensure the officer reads the entire report and then remove these intentionally added, erroneous details. Some of the examples of the erroneous entries in the user manual include:
The international spy network was revealed to be a group of highly trained squirrels.
The convicted criminal's sentence was to walk the plank in a community pool.
It was unanimously decided that rock-paper-scissors would determine the suspect's fate.
The criminal's hideout was inside a painting, accessed via a secret word.
Axon, 2025
On page 11 of the previous user manual (September, 2024 version), an obvious error insert within an example report can be seen reading as “A spell book was found at the scene, its pages fluttering open to a spell that turns metal to cheese…”, (awfully similar to the shape-shifting frog scenario from the news articles). The original article in the Park Record even discusses this obvious error feature, with the Heber City chief stating he was unaware of the setting (Taylor, 2025). Could it be this setting was turned on for West Jordan PD and neither his son-in-law who had the report with the frog details included, nor the Heber City Chief knew that feature was activated? If this were to be the case, it represents poor implementation of the product for these agencies based on a lack of training and understanding of functionality. It also brings into question the utility of this feature; human-in-the-loop is a must for all generative AI systems (especially when being used to expedite report writing), but inserting completely silly statements like this, which can obviously be missed by tired, over-worked cops is likely not the best way to ensure user review.
Problems with Certain AI Report Writing Solutions
If the former example I gave was the cause for the shape shifting frog details, it shows the inherent flaw in using body worn camera audio to generate reports. As any cop with a body worn camera knows, a large majority of the footage and audio on the camera is not relevant to an incident being investigated, especially if on a scene for an extended period of time. Generating reports from all of this audio is just not an efficient way to write reports and opens up the potential for instances like this to occur, based on irrelevant audio and transcription.
The news articles also discuss how the developers of Code Four, a competitive product to Axon, offer what they call a better solution, as their product uses both body camera audio and video to generate the report. Using video visualization with body worn camera footage opens up additional issues. As any law enforcement supervisor who has watched body worn camera video for a use of force review knows, camera angles do not capture everything an officer is seeing and processing in the live moment. This holds true for investigations as well. So, take for example, an officer is conducting Standardized Field Sobriety Tests on a subject in view of their body worn camera. Let’s say the officer is observing the walk and turn test and the footage captures the suspect stepping off the line or not touching heel to toe, but in the moment the officer does not see it. With video visualization writing the report off the footage, it would include those observation points in the report, even though it was not initially a determining factor for the officer to make an arrest as they missed it on scene, creating evidentiary issues.
The Answer? Policereports.ai
That’s why at Policereports.ai, we focus on allowing the officer to still control what is in his or her report from the start, either through audio dictation of the incident or uploading comprehensive written details. The officer is directly inputting the facts as he or she remembers, controlling the entire narrative of the report, just as he or she would if typing traditionally. Our system is customized for every client, formatting the report into the specific agency’s report structure and also completing agency specific forms, limiting manual data entry and duplicative work, increasing the efficiency and effectiveness of the officer.
Return here tomorrow to read part two of the post, which will discuss some reservations voiced by academia surrounding the use of artificial intelligence in public safety.
References
Axon Enterprise Inc. (2025). Draft One User Guide. In California Public Defenders Association. https://cpda.memberclicks.net/assets/MCLE-Events/axonwebinar/Axon%20Draft%20One%20User%20Guide.pdf
Daniel, L. (2026, January 4). Cop Transforms Into Frog, According To AI-Generated Police Report. Forbes. https://www.forbes.com/sites/larsdaniel/2026/01/04/cop-transforms-into-frog-according-to-ai-generated-police-report/
Taylor, C. (2025, December 16). Heber City Police Department test-pilots AI software. Park Record. https://www.parkrecord.com/2025/12/16/heber-city-police-department-test-pilots-ai-software/
To view my previous posts outlining our Policereports.ai product, follow these links below:
https://www.policereports.ai/blog/unpacking-ai-use-in-law-enforcement-document-completion
https://www.policereports.ai/blog/the-ai-implementation-paradox-bridging-the-genai-divide
