It is 11:47 p.m. on a Tuesday. Officer Ramirez has just cleared a domestic-disturbance call that ran ninety minutes: arrival, separation, two separate interviews, an EMS assist, a voluntary transport, and a supplemental neighborhood canvass. She is sitting in her cruiser with the in-car computer. The narrative is not going to write itself, and she still has four hours left on shift. In most agencies, the next ninety minutes belong to paperwork.

This lesson maps what happens next when AI is in the workflow. Not the version from a vendor slide deck, where the report appears fully formed and the officer taps "approve" while drinking coffee. The real version, where audio from a body-worn camera (BWC, the small camera clipped to an officer's chest or collar that records the encounter from the officer's point of view) is processed by a language model, a draft narrative appears in the records management system (RMS, the software platform that stores, organizes, and provides access to all agency records, from incident reports to arrest histories), and the officer reviews every word because the report is evidence, not a memo.

Chapter 3 of this program is the "AI across public safety today" tour: what the tools actually do right now, in 2026, at working agencies. Report writing is the natural first stop because it is where AI has made the most measurable operational difference and where the evidentiary stakes are highest. By the end of this lesson you will understand the mechanics of draft-from-footage, the performance data that makes the tool real, the failure modes that require human review to catch, and exactly where the line of human accountability sits.

The Paperwork Problem in Concrete Terms

Before getting to what AI does, it is worth understanding what it is being asked to replace. Law enforcement has a paperwork problem that is quantifiable and that consumes capacity that agencies cannot afford to waste.

Research consistently finds that patrol officers spend between 30 and 40 percent of every shift on documentation tasks: incident reports, arrest narratives, use-of-force documentation, follow-up supplements, CAD (computer-aided dispatch, the software platform dispatchers use to receive, log, and track emergency calls and unit deployment) entry corrections, and digital evidence logs. For a department running 10-hour shifts, that is 3 to 4 hours per officer per shift absorbed by typing. Across a 50-officer patrol division running three shifts, the math produces a staggering number of patrol hours that never reach the street.

The problem is not that officers are slow or inefficient writers. Police reports carry an unusual burden: they are legal documents, signed under oath, that must reconstruct a dynamic event with precision, in a specific administrative format, immediately after the adrenaline has cleared. Reconstructing a sequence of events accurately, in the correct order, with proper legal terminology for elements of each offense, takes cognitive energy and time even for experienced officers. For newer officers, the report queue becomes a source of stress that can follow them home and bring them in early on days off to catch up.

The question AI addresses is whether a machine can do the first pass on the narrative, freeing the officer to review and correct rather than compose from scratch. That is a fundamentally different and more tractable problem than asking AI to author a sworn document autonomously. Review and correction are faster than composition, they are less cognitively demanding, and they are the appropriate role for a tool that can generate a plausible narrative but cannot verify it against the record the way a witness to the event can.

Draft-from-Footage: How It Actually Works

The phrase "draft from footage" is the working description for the core capability: an AI system ingests the audio from one or more body-worn-camera recordings of an incident, transcribes it, identifies the incident narrative embedded in the audio, and produces a structured draft in the agency's report format. Axon's Draft One is the most deployed implementation of this capability in 2026; other vendors offer variations. The mechanics are worth understanding in detail because the failure modes are embedded in each step.

Step 1: Audio Capture and Upload

After an incident, the officer docks the BWC or initiates a wireless upload. The footage, including its embedded audio track, moves to the agency's cloud evidence platform. This upload is governed by the agency's CJIS (Criminal Justice Information Services, the FBI program that sets the security policy for all law enforcement data systems, including required encryption, access controls, and data handling standards) security requirements. The agency, not the vendor, is responsible for CJIS compliance; vendor contracts may include CJIS-compliant hosting, but the obligation of compliance stays with the law enforcement agency.

At this step, the failure risk is straightforward: audio quality. BWC microphones vary. Wind, crowd noise, and distance all degrade transcription accuracy. A partial transcript produces a partial draft. The officer reviewing the draft must be alert to sections where the audio was unclear, because those sections are where the model is most likely to fill gaps with inferred detail rather than transcribed reality.

Step 2: Transcription and Narrative Extraction

The uploaded audio is processed by a speech-to-text model, which converts the spoken audio to a written transcript. Then a second pass, using a large language model (a generative AI system trained on large text corpora), identifies the narrative elements relevant to a police report: incident type, involved parties, location, observed facts, statements made by parties, officer actions, and disposition. This is not a simple transcription copy-paste. The model is doing narrative extraction, identifying which parts of a ninety-minute recording are report-relevant and organizing them into a draft structure.

This step is where the most important failure modes live. Transcription errors can produce a word the officer never said. Narrative extraction errors can place events in the wrong sequence or include a detail from a background conversation that does not belong in the narrative. And the model, trained on police report patterns, will sometimes infer detail that fits the incident type based on training data rather than what the footage actually shows. That inferred detail is not marked or flagged; it appears in the draft alongside accurately transcribed content.

The draft narrative is a hypothesis about what happened, drawn from audio evidence. The officer is the only person in a position to confirm or refute it.

Step 3: Draft in the RMS

The extracted narrative is populated into a draft report in the agency's RMS. Depending on the system, the draft appears in the officer's report queue as a pre-populated incident report, or it is presented in a side-by-side interface where the draft and the BWC footage can be reviewed simultaneously. Some platforms timestamp the draft text so the officer can jump to the corresponding footage segment for each sentence.

This is the interface where the officer does the work that makes the report legally sound. The draft is a starting point. The officer's job is to verify every factual claim in the draft against the actual footage, correct any inaccuracies, fill in context that the audio did not capture (observations visible in the footage that were not narrated, physical evidence, conditions on scene), and adopt the narrative as their own sworn account.

Step 4: Officer Review and Adoption

This step is not optional, and it is not a formality. Adoption of the AI draft as a sworn report is a legal act. The officer who submits the report is representing, under oath, that the narrative accurately describes the incident. That representation does not change because the first draft was generated by software; it stays with the officer regardless of how the narrative was produced.

Agencies that have deployed draft-from-footage tools have learned that officer review discipline varies. Some officers do the full footage-grounded verification pass, catching errors and adding missing context. Others read the draft quickly and submit it with minimal review. The difference matters enormously: a report that was carefully verified against the footage and an error caught is a defensible sworn document. A report that was AI-drafted and submitted without adequate review is a liability waiting to surface in discovery.

The 82 Percent Finding and What It Actually Means

Axon's testing of Draft One found that officers using the tool reported an 82 percent decrease in report-writing time. This is the most widely cited performance figure for AI report writing in law enforcement, and it deserves careful interpretation because both its significance and its limits are real.

The significance is clear. If an officer who previously spent 90 minutes on a narrative can produce a verified report in under 20 minutes, the math is transformative for patrol capacity. Across a shift and a department, those recaptured minutes represent real return-to-patrol time, faster response, and cognitive bandwidth that can go toward the parts of the job that require presence and judgment. The figure is not just a marketing claim; it represents a genuine change in the economics of shift work.

The limits require equal attention. First, the 82 percent figure comes from officer self-reporting during a testing phase, not from a controlled longitudinal study. Officers evaluating a new tool may report time savings that are somewhat inflated by the novelty effect, the change in routine, and the motivation to make the tool work. Second, the time savings assume adequate review. An officer who runs through the draft in 5 minutes without checking it against the footage may save time compared to composing a narrative from scratch, but the resulting report has not been verified. Time saved on an unverified report is not a win; it is a deferred problem. Third, the savings vary by incident type. Simple, well-recorded incidents with clean audio produce drafts that are faster to verify. Complex incidents, those with poor audio, multiple parties, disputed facts, or use of force, produce drafts that require more intensive review and may be nearly as time-consuming as hand-composing the narrative.

The honest framing is this: AI report drafting returns significant time to patrol when used with proper review discipline. The time savings and the evidentiary quality are not in tension if the review is done right. They are only in tension when the tool is treated as a shortcut rather than as a starting point.

The Specific Failure Modes That Review Must Catch

Earlier lessons in this chapter introduced the general category of AI hallucinations. In the report-writing context, there are three specific failure modes that officers reviewing AI drafts need to be actively looking for.

The Gap-Fill Detail

When the audio for an incident is incomplete, unclear, or missing for a portion of the event, the model does not produce a gap with a placeholder. It generates text that fits the narrative. A model that has been trained on thousands of domestic-disturbance reports knows what domestic-disturbance reports typically contain. If the audio does not capture whether the officer observed signs of physical struggle in the residence, the model may write that the officer "observed disarray consistent with a physical altercation" because that is what typically appears at that point in a narrative of that incident type.

That sentence, if it appears in the draft and the officer does not catch it, becomes part of the sworn narrative. If the officer did not actually observe disarray, or if the disarray is disputed at trial, the defense attorney will ask the officer to describe exactly what they observed. The footage will not show what the report claims. That is a credibility problem, a Brady (Brady v. Maryland, the 1963 Supreme Court case requiring the prosecution to disclose exculpatory evidence to the defense) problem, and potentially a basis for suppression.

The gap-fill detail is hard to catch precisely because it is plausible. It fits. It sounds like something an officer would write. The only way to catch it is to read every factual claim in the draft and ask: "Is this in the footage, or is this inferred?" That question requires actually reviewing the footage, not just reading the draft.

The Softened Fact

Language models optimize for fluent, coherent prose. In a police report context, that optimization can produce language that softens the description of an event. An officer who said on camera that the subject "lunged toward me aggressively" may find the draft reads "moved in the officer's direction." The event is not wrong in the draft, but the characterization is weaker. In a use-of-force case, that difference matters: the legal standard for use of force depends on the officer's reasonable perception of threat, and the language of the narrative is part of the evidentiary record of that perception.

Softened facts are also easy to miss in a quick review because they do not produce a factual mismatch. The reviewer is looking for errors, and the softened version is not technically an error. It is an insufficiency, and catching it requires comparing the draft language against the officer's actual memory of the event and the footage.

The Invented Quote

Direct quotes are among the most legally consequential elements of a police report. A subject's statement, a witness's account, a Miranda (the advisement of constitutional rights required before custodial interrogation, established in Miranda v. Arizona) advisement acknowledgment, all of these carry specific evidentiary weight and are subject to direct scrutiny at trial. A quote in the report must be what the person actually said, to the degree the officer can reconstruct it from memory and the recording.

AI models are capable of generating plausible-sounding quotes that no one actually said. If the audio captured a subject making a statement but the words were partially inaudible, the model may "complete" the quote with language that fits the context. That completed quote is fabricated evidence. At trial, the defense can play the audio and demonstrate that the words attributed to the subject in the report were never spoken. The implications for the case and for the officer's credibility are severe.

The rule on quotes is absolute: every quoted statement in a report must be verified against the audio. If the audio is unclear, the report must reflect that uncertainty: "Subject stated words to the effect of..." or a description of the substance of the statement rather than a direct quote. A fabricated quote is worse than no quote.

The King County Line and What Disclosure Looks Like

In 2024, the King County Prosecuting Attorney's Office in Washington State announced a policy barring AI-written police reports from its prosecutions unless the AI involvement was clearly documented and the human review was verified. This is the most significant prosecutorial governance line on AI report writing in the country as of 2026, and it represents a challenge that agencies using draft-from-footage tools need to answer directly rather than avoid.

The King County position was not that AI is bad. It was that an undisclosed, unverified AI draft creates a discovery problem. The defense has a right to know how the report was produced, and the prosecution needs to be confident that the report reflects the officer's verified account rather than an AI model's narrative reconstruction. The concern was not about AI per se; it was about the verification standard and the documentation of that standard.

The Electronic Frontier Foundation (EFF), the digital-rights advocacy organization, has raised parallel transparency concerns: that AI-generated reports without disclosure deprive defendants of information relevant to challenging the account, and that the opacity of AI involvement in evidence creation is itself a constitutional question. The EFF's position deserves to be understood on its own terms, not dismissed as activist opposition to technology. The question "how was this report produced?" is a legitimate question for a defense attorney to ask, and the answer needs to be documented and defensible.

What disclosure looks like in practice depends on the agency and the prosecutor. At a minimum, most guidance suggests that the report file note whether an AI tool was used to assist with drafting, that the officer's review and adoption be documented, and that the footage verification be logged. Some agencies are developing specific disclosure language for the report header or for a metadata field in the RMS. The principle is transparency by design: if the question comes up in discovery or at trial, the officer and the agency have a clear, documented answer, not an awkward silence.

Bundled Contracts and the Lock-In Question

One dimension of AI report writing that agencies and officers need to understand, even if it feels more like a policy question than a patrol question, is the procurement structure that delivers these tools. Axon and similar vendors have moved to bundled, multi-year, sole-vendor contracts that package body-worn cameras, digital evidence storage, drones, tasers, and AI services including Draft One into a single agreement. Contracts of approximately $45 million and up to 10 years in duration are documented in the procurement record as of 2026.

The operational significance for working officers is real, even if procurement feels remote from the patrol car. When an agency signs a 10-year bundled contract, the AI report-writing tool embedded in that contract is not going anywhere for 10 years. The agency's ability to switch to a different vendor's tool if the AI quality is inadequate, if the pricing becomes unreasonable, or if the tool produces systemic accuracy problems is severely limited. The officer using Draft One in year three of a 10-year contract is using it regardless of whether a better tool has emerged, and the agency's negotiating position with the vendor is constrained by the same contract.

This does not mean bundled contracts are wrong. There are legitimate efficiency and interoperability arguments for single-vendor platforms. But it does mean that the verification and governance standards agencies build around AI tools need to be written into policy at the time of procurement, before the contract locks in the tool. An agency that signs a 10-year AI report-writing contract without a verified review standard is committed to 10 years of uncertain evidentiary quality. An agency that signs the same contract with a verified review standard is committed to 10 years of a well-governed tool. The difference is entirely in the policy, not the technology.

Key Takeaways

• Officers currently spend 30 to 40 percent of every shift on documentation. AI draft-from-footage tools address this by generating a first-pass narrative from body-worn-camera audio, allowing the officer to review and correct rather than compose from scratch.
• Axon Draft One testing found officers reported an 82 percent decrease in report-writing time. That figure assumes proper review discipline; time saved on an unverified report is not a gain, it is a deferred liability.
• The three specific failure modes in AI report drafts that review must catch are the gap-fill detail (the model invents a fact to fill an audio gap), the softened fact (the model smooths the language and weakens the characterization), and the invented quote (the model completes a partial or inaudible statement). Each is a potential Brady problem and a credibility risk at trial.
• Adoption of an AI draft as a sworn report is a legal act. The officer's review must verify every factual claim against the footage. "The computer wrote it" is not an acceptable answer in a deposition; authorship and responsibility stay with the officer regardless of how the draft was produced.
• The King County prosecutorial policy and EFF's transparency concerns both converge on the same principle: disclosure of AI involvement in report drafting is required, not optional. The agency needs a documented, defensible answer to the question "how was this report produced?" before the question is asked in court.
• CJIS Security Policy obligations for data handling stay with the agency, not the vendor. Bundled, multi-year contracts of approximately $45 million and up to 10 years in duration are now common; governance and review standards need to be built into policy at procurement, before the contract locks in the tool for a decade.
• The appropriate framing for AI in report writing is not "AI writes the report." It is "AI produces the starting point, and the officer produces the report." That distinction is not semantic; it is the difference between a defensible sworn document and an AI output that has not been verified against the record.