AI-Generated Police Reports Don’t Work the Way You Probably Think

Axon’s Draft One does not simply “watch” body-camera video and write a report. Understanding what the system actually receives, transforms, and generates matters to understanding the evidence it produces.

Take Axon’s Draft One.

When Draft One launched in 2024, agencies could use it in two ways: those using Axon Records could generate Draft One narratives within Axon’s records-management system, while those using another records-management system could generate the narrative through Axon Evidence and copy it into their own system.

The architecture has evolved. Axon now makes a Report Writer interface available within Axon Evidence as well as within its own records-management products. For agencies using a third-party records-management system, Report Writer can serve as the workspace for tools like dictation and Draft One before the resulting narrative is copied into the agency's own system. For those within Axon's records-management ecosystem, that same Report Writer environment allows an officer to type or dictate a report, import an agency-configured template, or invoke Draft One to generate a draft.

In other words, Draft One is increasingly one tool inside a broader Axon report-writing environment rather than a standalone report-writing workflow. And Draft One itself no longer requires the officer to begin with body-camera evidence; it can also generate from typed or narrated incident details.

When Draft One is used with body-camera recordings, however, something important happens. The system displays recent recordings as “evidence” files. The officer checks the box next to the recording they want Draft One to use. But Draft One is not actually analyzing those videos as videos. Behind the interface, it retrieves the transcripts associated with the selected recordings. It does not use computer vision to analyze what the camera sees.

At "Step 1," Draft One asks officers to select body-camera recordings as "evidence" for report generation. When a recording is selected, Draft One uses the associated audio transcript rather than analyzing the visual content of the video. Source: Axon.

Instead, Axon’s Auto-Transcribe system converts the body-camera audio into text, using speaker labeling to try to distinguish speakers amid overlapping dialogue and chaotic noise. Additionally, for agencies using Axon's Records or Standards platforms, beginning in April 2026, Draft One also began using information already entered into report fields, such as people's names or vehicle information, in generating the draft. Axon encourages officers to complete the report fields before generating the narrative.

Once the officer has selected the “evidence” to base the narrative on, Draft One walks the officer through a series of structured inputs that, together, build the prompt the model will use to generate the report.

First, Draft One presents a checkbox menu asking the officer to characterize the incident: incident type, charge severity, whether an arrest was made, and desired draft length. Those selections do not look like a prompt in the familiar ChatGPT sense, but they are constrained instructions that shape the task the model is being asked to perform:

At Step 2, Draft One asks the officer to characterize the incident and choose the desired length of the generated narrative. Source: Axon.

Also beginning in April 2026, after those selections, Draft One analyzes the selected evidence transcripts and presents AI-generated prompts for additional information. Axon says these prompts are designed to elicit missing details. The officer can answer those questions, view additional suggestions, switch to a broader narration interface, or skip this step entirely.

At Step 3, Draft One presents suggested questions based on the information and evidence already provided, inviting the officer to supply additional details before the draft is generated. Source: Axon.

If the officer switches to the narration interface, they are prompted to describe the call, arrival, people involved, events, decisions, evidence, and outcome, and are also provided a free-text field for typing or narration.

Draft One also allows officers to type or narrate additional information before generation. The interface tells the officer that Draft One will work those details into the resulting narrative. Source: Axon.

There is another consequential feature of this workflow. Axon says that the narration and typed text supplied at this stage are not retained in the system or audit logs. The audit trail records that those inputs were provided, but not what the officer actually said or typed. In other words, information can enter the generation pipeline, shape the official police narrative, and then be destroyed rather than preserved for later review.

For lawyers, that raises an immediate discovery and impeachment problem. In many jurisdictions, officer statements are independently discoverable under statutes, rules, or local discovery regimes. Further, if the officer's statements materially differ from the generated report or from later testimony, those differences could be significant enough to trigger constitutional disclosure requirements as impeachment material. But if the underlying input has already been destroyed, neither prosecutors nor defense counsel can later examine that account or review any discrepancies.

Like the prompt you write in Claude or ChatGPT, those inputs (including the checkbox selections, answers to Draft One's suggested questions, and additional context typed or narrated by the officer) help structure the model's task and determine its output. Only after those structured prompting steps are completed or skipped does the large language model generate the draft.

In "AI Is Not One Thing," I argued that we need to stop talking about “AI” as though it were one thing. Draft One shows why.

This “AI-generated police report” workflow combines speech-to-text transcription, speaker labeling, and a large language model. It is run through a user interface that determines what information reaches the model and through human choices that structure the model’s task before it generates a single sentence.

But the architecture does something else too. It changes who can shape the record.

Axon knows that this architecture can change officer behavior. Its own best-practices guidance tells officers to narrate key facts aloud, echo back what community members say, and ask questions on scene because more dialogue in the transcript helps Draft One generate a fuller report. And the company says officers have reported doing exactly that. The people on the other side of the camera generally do not have that same knowledge.

That creates an asymmetric feedback loop.

Think about a familiar phrase: “Stop resisting.” Sometimes the words and the images tell very different stories. In a system built around audio transcription, the words “stop resisting” are available to the report-generating model. What the video shows at that same moment is not. The separation is built into the workflow. Axon says the transcript can be available to Draft One within minutes of a recording ending, while the full video may not reach Axon Evidence until the camera is docked. That does not mean Draft One will necessarily say the person resisted. It means something more basic: one part of the evidentiary record enters the generation pipeline while another does not.

That matters because many criminal cases begin—and end—with the police report. Charging and plea decisions can develop around that narrative before anyone meaningfully scrutinizes the body-camera footage.

Draft One is not simply “writing reports faster.” It changes behavior, filters evidence, structures interpretation, and helps produce the narrative that enters the legal system.

To understand an “AI-generated police report,” we have to examine that entire chain—and ask not only what the technology can see, but who decides what it gets to see.