Officer Martinez finishes her third call of the morning, a welfare check that turned into a two-hour mental-health response, and returns to the station to find her report queue has four incidents waiting. Across the building, a dispatcher is fielding a call on three lines at once while her computer-aided dispatch (CAD) system flags a potential duplicate address. In the records unit, a clerk is staring at a stack of body-worn camera (BWC) footage release requests, each one requiring individual review and redaction before it ships out the door. All three of them have heard something about AI in the past six months. All three have a different picture of what that word means. The officer thinks of a humanoid robot from a sci-fi movie. The dispatcher thinks of the chatbot she tried on her phone. The records clerk isn't sure what she thinks, but she's heard the word "hallucination" in a training and it made her nervous. This lesson is about replacing all three pictures with an accurate one, because a tool you misunderstand is either a tool you fear without reason or a tool you trust without caution, and neither is safe in public safety work.

The Three Kinds of AI in Public Safety Right Now

Before we discuss what AI can do in a patrol car, a dispatch console, or a records unit, we need to sort the term itself. "AI" is used to mean at least a dozen different things in the news, in vendor pitches, and in policy debates. For working public safety professionals, there are three categories worth understanding because they show up in three distinct places in your workday, they work in completely different ways, and they fail in completely different ways.

Transcription and Summarization

The first category is transcription and summarization. This is the kind of AI that listens to audio, converts it to text, and can produce a condensed summary of what was said. If you have used a voice-to-text tool on your phone, you have experienced a basic version of transcription AI. In a public safety context, transcription AI can take the audio track from a body-worn camera recording and produce a text transcript of what was said during an encounter. It can take the audio from a recorded interview and produce a text of the interview. It can take a recorded 911 call and produce a readable log.

Transcription AI is enormously useful for reducing the hours a detective spends listening and re-listening to interviews. A two-hour interview that used to require two hours of listening time for every pass a detective took through it can now be skimmed as text in fifteen minutes. That is a genuine time return, and it is happening in agencies right now. But transcription AI is not perfect: it mishears words, it conflates similar-sounding names, it struggles with accents and background noise, and it occasionally drops a sentence or reverses the order of two statements. Every transcription used in a legal context needs human review against the original recording. This point will come up many times across this program because it is the theme that holds everything together: the AI output is a draft, and a human with access to the source material is the verification step.

Summarization AI takes a body of text, whether a transcript, a case file, a collection of records, or a set of reports, and produces a shorter version that captures the key points. Summarization is useful when a detective needs to orient to a case with hundreds of pages of prior documentation. It is useful when a records clerk needs to understand the scope of a public records release request before diving into the file. And it is useful, carefully used, when a supervisor needs to get up to speed on an incident quickly. But summarization AI compresses, and compression always involves choices about what to keep and what to leave out. A summarization that leaves out an exculpatory detail is not just a bad summary; in some contexts, it is a problem with constitutional implications. We will spend significant time on this in later lessons.

Generative Drafting

The second category is generative drafting. This is the most powerful and the most legally significant kind of AI in public safety today. Generative AI does not just transcribe what was said or compress what was written. It produces new text. It generates prose. It drafts a narrative. And the specific application that is transforming patrol work in 2026 is report drafting from body-camera footage.

Axon's Draft One, the product that is currently deployed at multiple agencies and bundled into multi-year camera contracts, drafts a police report narrative from the audio of the body-worn camera recording. The officer activates the camera, works the call, and when the encounter is over, the system has already processed the audio and produced a draft narrative. Officers who tested the product in pilot programs reported an 82% decrease in the time they spent on report writing. Because an officer currently spends roughly 30 to 40% of every shift on paperwork, that 82% time reduction translates to roughly 25 to 30 percentage points of shift time returned to patrol, community engagement, and the actual work of policing. That is not a small number. It is a meaningful operational change.

But here is what generative drafting actually does, technically, and why this matters for every officer who uses it. The system is not playing back a transcript of the encounter and organizing it into report format. It is generating text based on the audio, the language patterns it was trained on, and its statistical model of what a police report narrative for this kind of incident looks like. That process is extraordinarily fluent. The prose it produces usually sounds exactly like a well-written police report. But "sounds like" is not the same as "accurately reflects." The model can generate a detail that is consistent with the audio but was not actually said. It can construct a sequence of events that makes narrative sense but differs from the actual sequence on the recording. It can fill in a gap in the audio, a moment where the recording was unclear or the officer moved out of range, with something that sounds plausible but is invented.

This is not a flaw the manufacturer forgot to fix. It is how generative language models work. They generate the next most probable piece of text given everything they know. They do not know the difference between text they generated from something that actually happened and text they generated to fill a gap. Both come out with the same confident, professional tone. This is the fact that reorders every rule about AI use in public safety, and we will return to it throughout this program.

Classification and Prediction

The third category is classification and prediction. This kind of AI looks at data, usually structured data with many attributes, and assigns a category or a score. In public safety, classification AI appears in several places. It can suggest a call type to a dispatcher based on the text of the caller's description. It can flag a record as potentially requiring redaction before human review. It can classify an incident report into a category for statistical analysis. Predictive AI can estimate the likelihood that a particular address will generate another call within a given time window.

Classification and prediction AI is neither as dramatic as the generative kind nor as immediately transformative. But it is deeply embedded in the systems public safety agencies already use, often without the agency being fully aware of it. Modern CAD systems include classification assistance. Records management system (RMS) search tools include relevance ranking that is, technically, a form of prediction. Understanding that these are AI functions, not neutral mechanical processes, matters because classification AI can be wrong, and it can be wrong in patterns, and patterns of errors in public safety classification have legal and civil-rights implications.

Killing the Robocop Framing

The most important misconception to dispel early in this program is the "robocop" picture of AI: the idea that AI in public safety means autonomous decision-making, a machine that replaces human judgment, an electronic officer that decides who to stop, who to arrest, or how to classify a call without a human in the loop. That picture is not just wrong; it is dangerous, because it points the wrong direction for both the opportunities and the risks.

None of the AI currently deployed in public safety agencies operates autonomously on life-safety decisions. Generative report-drafting AI produces a draft. A human reads it, corrects it, and adopts it. Transcription AI produces a text. A human verifies it against the recording. Classification AI suggests a call type. A human dispatcher owns the priority. These are tools that assist human decision-making and human authorship. They do not replace them, and no responsible deployment would have them do so.

The robocop framing is dangerous not because it overstates the power of current AI but because it misdirects attention. If you are worried about AI replacing human judgment, you may not notice the more immediate and real risk: that AI-assisted human judgment can produce errors that are harder to detect than unassisted human errors, because the errors come pre-packaged in fluent, confident prose that reads like it was carefully written by an experienced officer. A detective who writes a bad summary is visibly authoring a bad summary. An officer who adopts an AI draft with a hallucinated detail is, in many cases, unaware that the detail is not in the footage, because the prose is so fluent that it does not prompt the kind of skeptical re-reading that a clumsy human draft would invite.

The risks in public-safety AI today are not about autonomous machines making decisions. They are about subtle errors in text that looks authoritative, about disclosure obligations when a new kind of tool is in the chain of evidence production, and about accountability when something goes wrong and someone asks who wrote this. Those are the risks this program addresses.

AI in public safety today is a powerful drafting and summarization tool held in a human author's hand. The human author is accountable. The tool is not.

Where AI Actually Sits in the Patrol Workday

Let's put the three categories of AI into the specific places they appear in a working patrol officer's day, because abstraction is harder to act on than a concrete map.

Before the Call

Before the officer even arrives at a scene, the CAD system that dispatched her may have used classification AI to suggest a call type and priority level based on the dispatcher's typed summary of the caller's description. This is happening in the background, usually invisibly. The dispatcher sees a suggested priority, may or may not actively notice it, and either accepts it or overrides it with their own judgment. Most of the time the suggestion is consistent with what the dispatcher would have chosen independently, so it is not noticed. When it is wrong, the consequences range from a minor inefficiency (the wrong number of units dispatched) to a serious outcome (a call misclassified as lower priority than the situation warranted).

The Records Management System (RMS), which connects to the CAD, may also use AI to surface related prior incidents, known addresses, or relevant records when the dispatcher or officer pulls up a location. This is AI-assisted information retrieval. It is useful for situational awareness, but the returned records are ranked by relevance estimates, not by certainty. A record that appears near the top of the results is ranked high because the system estimated it was likely relevant, not because a human reviewed the current call and confirmed the relevance.

During the Encounter

During the encounter, the body-worn camera is recording. In agencies using modern camera platforms, the footage is automatically uploaded to a cloud evidence platform. In agencies using Draft One or a comparable product, the audio is being processed in the background during or immediately after the encounter. By the time the officer returns to the station or sits down in the patrol car with a laptop, a draft narrative may already be available.

This is where the workflow transformation is most visible. The officer who used to sit down at the RMS terminal and spend forty-five minutes or an hour writing a narrative from memory and notes now sits down to a draft that covers the key facts of the incident in the structure of a police report. The time saving is real. The opportunity for error is also real, and it is concentrated in a specific place: any fact, any sequence, any detail in that draft that is not present on the recording or in the officer's own notes is a potential hallucination.

The verification discipline that turns a useful AI tool into an acceptable evidence-production tool is to read every sentence of the draft against the footage and notes and ask, for each factual claim: is this in the recording? Can I point to the moment? This is not a casual read-through. It is a sentence-level accountability pass. That is the standard, and it is not optional. It is optional in the same sense that signing a report is optional: technically, nothing stops you from skipping it. But the professional and legal consequences of an AI-generated error in a sworn narrative are the consequences of the officer who adopted the draft as their own report.

After the Call: Records and Release

After the call is documented, the footage and the report enter the records lifecycle. At some point, there will be a public records request, a discovery request from a defense attorney, or an internal review. AI enters this part of the workflow as a redaction tool. AI-assisted redaction can identify faces, license plates, addresses, and other personal identifiers in footage automatically and propose them for redaction before the footage is released.

Redaction AI genuinely helps with the workload problem. A records unit receiving a public records request for body-camera footage from an incident involving bystanders faces a review-and-redaction task that can take hours per video. AI-assisted redaction can complete an initial pass in a fraction of the time, flagging the likely redaction areas for a human reviewer to confirm. The risk is the missed identification: a face that the AI did not recognize, a partial plate that was not flagged, or a name that appeared on a document in the background and was not caught. Over-redaction (blurring too much) and under-redaction (missing something) are both failures, and both can produce legal exposure.

The Dispatch Console: A Closer Look

The telecommunications center, where dispatchers work, deserves its own section because AI appears there in ways that are both more continuous and less visible than in patrol work.

A modern 911 communications center processes hundreds or thousands of calls per day. Every call generates a CAD entry, a text record of the incident that drives the dispatch of resources. AI assists this workflow in three specific places. First, real-time transcription of the caller's words can produce a running text display alongside the audio, helping the dispatcher track what is being said in high-stress, noisy, or linguistically complex calls. Second, automated translation AI can provide approximate English text of calls in other languages, giving the dispatcher a starting point while a live interpreter is connected. Third, classification AI, as discussed above, can suggest a call type and priority based on the content of the transcribed caller description.

In all three cases, the AI output is assisting the dispatcher, not replacing them. The dispatcher hears the caller directly, assesses the situation with the full context of everything they know about the area, the time of day, the prior call history, and the current resource picture, and makes the priority and dispatch decisions. The AI outputs are inputs to that judgment, not substitutes for it. The risk in the dispatch environment is the same as it is everywhere else in AI-assisted public safety: the confident-looking AI suggestion that is wrong, presented in the same confident format as a correct suggestion. A dispatcher who has learned to trust AI classification suggestions without actively reviewing them will occasionally dispatch based on a wrong classification, and in dispatch, a wrong classification can mean a delayed response to a life-threatening situation.

The Records Unit: Context and Scale

The records unit operates at a different pace than patrol or dispatch, but the AI challenges are in some ways more concentrated there because the outputs travel further and carry more legal weight. A public records release from a police department can go to a journalist, a civil litigator, or a community member, and any error in that release (over-redaction, under-redaction, or incorrect categorization of releasable versus exempt material) can generate a legal challenge, a complaint, or a news story.

AI appears in records work in three main places. First, AI-assisted redaction of body camera footage, as discussed above. Second, AI-assisted review of written records for personal identifying information before release under open-records statutes. Third, in more advanced deployments, AI-assisted classification of records by exemption category, helping records staff make initial decisions about what portions of a file are covered by law enforcement exemptions, privacy exemptions, or active-investigation holds.

Each of these applications saves time. Each carries the same underlying risk: the AI that is confident but wrong, the missed item, the incorrect classification. And the consequences are amplified in the records context because the outputs are reviewed by people outside the agency, people with legal training, journalists, and civil-rights advocates who are actively looking for errors. The Electronic Frontier Foundation (EFF) and similar organizations have raised specific concerns about AI use in police records systems, particularly around whether AI-assisted redaction decisions are being adequately reviewed by humans before release. These concerns are legitimate and worth understanding on their own terms, not just as regulatory compliance issues. An agency that releases footage with an AI redaction miss is an agency that has failed a real person whose privacy was not protected as the law requires.

Key Takeaways

• AI in public safety today falls into three distinct categories: transcription and summarization (converting audio to text and condensing it), generative drafting (producing new prose from recordings and data), and classification and prediction (assigning categories or scores to data). Each category works differently, fails differently, and requires different verification habits.
• Generative drafting AI, including products like Axon's Draft One, drafts police report narratives from body-camera audio. Officers in pilot programs reported an 82% decrease in report-writing time, representing a major operational shift. Officers currently spend roughly 30 to 40% of their shift on paperwork, so this time return is significant.
• The "robocop" framing, AI as an autonomous decision-maker that replaces human judgment, is inaccurate for current public-safety AI deployments. AI in the field today assists human decision-making; it does not replace it. The real risk is not autonomous AI but subtle errors embedded in fluent, confident-sounding AI output.
• The CAD (computer-aided dispatch) system, the RMS (records management system), the body-camera evidence platform, and the records release workflow all contain AI functions, often invisibly. Understanding where AI sits in the workday is the first step toward managing it responsibly.
• AI outputs in all three categories require human verification against source material: the recording, the CAD entry, the original document. A detail in an AI draft that cannot be traced to the source is a potential hallucination and must not be adopted into a sworn narrative.
• The EFF and prosecutor offices like King County, Washington have raised legitimate governance concerns about AI-assisted police reports and records. An officer who understands both the opportunity and the concern is an officer who can operate the tool responsibly and explain that operation to a prosecutor, a defense attorney, or an oversight board.
• The accountability for every AI-assisted output in public safety stays with the human who reviews and adopts it. The officer who adopts an AI report draft is the author of that report. The dispatcher who acts on an AI classification suggestion owns that dispatch decision. The records clerk who approves an AI redaction pass owns that release. The tool assists; the professional is responsible.
• CJIS, the Criminal Justice Information Services Security Policy, governs how criminal justice data is handled. When AI systems process that data, the agency's CJIS obligations do not transfer to the vendor. Understanding where your data goes when an AI system processes it is a compliance requirement, not an optional detail.