Training Coordinator Sandra Okafor had three rooms to fill and exactly forty-five minutes in each. Patrol officers in the first room spent their shifts writing reports and working calls. Telecommunicators in the second room sat at computer-aided dispatch (CAD, the system that receives 911 calls, assigns units, and logs incident data) consoles and handled emergency calls in real time, sometimes managing six open incidents simultaneously. Records clerks in the third room processed case files, fulfilled public-records requests, and maintained the records management system (RMS, the agency's database for incident, arrest, and case records). All three groups were being trained on the same AI rollout, and all three groups needed a fundamentally different forty-five minutes.

Why Role-Specific Training Is Not Optional

The default in most agency AI rollouts is a single training session for everyone, delivered by a vendor trainer, focused on the mechanics of the tool: how to log in, how to generate a draft, how to submit a report. That session leaves everyone with the same gap: they know how to use the tool but not why the specific verification and disclosure disciplines their role requires are built the way they are.

The verification standard for a patrol officer running a footage-grounded verification pass on an AI-drafted use-of-force report is not the same verification standard as a telecommunicator checking an AI-assisted call summary against the call recording before it enters the CAD. Both standards exist. Both are necessary. They are not interchangeable, and a training session that treats them as the same creates officers who apply the wrong standard in the wrong context, which is frequently worse than no standard at all.

Sandra's three rooms were not a logistical nuisance. They were the correct structure for a curriculum that takes the evidentiary environment of each role seriously. The patrol officer's AI output is a sworn report that will be disclosed to the defense, read in depositions, and tested at trial. The telecommunicator's AI output is a CAD entry that determines dispatch priority and becomes part of the incident record. The records clerk's AI output is a redacted case file released under open-government statutes, where a missed face or license plate is a privacy violation and a potential lawsuit. Each is evidence. The verification standard that is sufficient for one is not sufficient for another.

Patrol: The Report-Writing Curriculum

For patrol officers, the curriculum has a single organizing principle: the officer is the author of the sworn report, and the AI is a drafting tool that requires verification before the report can be submitted. That principle sounds simple. Making it operational requires training in four distinct competencies.

Competency One: Recognizing Gap-Fills

A gap-fill is what happens when the language model generating a report from BWC audio encounters a moment the microphone did not clearly capture and fills it with language that is statistically likely to be accurate based on the call type and patterns in its training data. Gap-fills are not random. They follow predictable signatures: specificity without audio support (a detail the officer never narrated), use-of-force boilerplate (phrases like "threatening manner" that appear frequently in training data), and scene descriptions without an audio source (vivid environmental details the microphone could not have generated).

Patrol training should include a gap-fill identification exercise. Take five AI-drafted reports from the pilot cohort (with identifying information removed) and run a gap-fill hunt as a group. The exercise is not to grade officers on finding every gap-fill. It is to build the recognition pattern in people who have never systematically looked for it. Officers who have done this exercise once are significantly better at catching gap-fills in their own drafts than officers who have only been told gap-fills exist.

Competency Two: Running the Footage-Grounded Verification Pass

The footage-grounded verification pass is the core skill of the program. It involves opening the AI draft and the BWC footage simultaneously, working through the draft paragraph by paragraph, identifying every factual claim, and confirming each claim against the footage at the relevant timestamp. Claims that cannot be traced to the footage, the CAD entry, or a specifically documented personal observation are ungrounded and must be corrected before submission.

Training should include a live demonstration with a real BWC recording and a real AI-drafted report. The trainer runs the pass in front of the group, narrating every decision: "This claim says the subject was facing north. I'm going to timestamp 4:17 in the footage. The footage shows the subject is facing northeast. I'm correcting the claim." That level of explicitness is uncomfortable for trainers who are used to talking about process in the abstract. It is the only way to make the pass concrete enough that officers can replicate it independently.

After the demonstration, officers should run the pass on their own, using a training dataset. The training coordinator should build that dataset in advance: five to ten BWC recordings with pre-generated AI drafts that contain known errors. Officers complete the pass, document their findings, and compare results. The comparison is the learning moment: officers who missed a gap-fill understand why they missed it because they can see it in the footage side by side with the draft language.

Competency Three: Disclosure Language and the Audit Log

Every AI-assisted report submitted by a patrol officer should include a disclosure note. The disclosure note does two things: it meets the Brady v. Maryland and Giglio v. United States disclosure requirements by making the AI assistance transparent, and it clearly attributes the sworn account to the human officer who verified and adopted it.

Training should provide officers with the agency's approved disclosure language, explain why each element of the language is there, and have officers practice inserting it into draft reports. Officers who understand why the disclosure language exists are more likely to use it consistently than officers who are told it is required policy. The why is not complicated: a prosecutor who knows the report was AI-assisted and verified can answer defense counsel's question. A prosecutor who does not know that cannot.

The audit log is the companion to the disclosure note. The log records when the verification pass was run, what claims were checked, what corrections were made, and when the final report was submitted. Officers should be trained to maintain this log in whatever system the agency designates, whether that is a field in the RMS, a note in the BWC platform, or a separate log maintained by the supervisor. The format matters less than the consistency. Every AI-assisted report should have a corresponding log entry that can be produced in discovery.

Competency Four: The Deposition Answer

The deposition question is coming. "Officer, did you write this report, or did a computer?" Every patrol officer trained on AI-assisted reporting needs to know how to answer it. The answer is not "the computer wrote a draft and I reviewed it." The answer is: "I am the author of this report. An AI tool produced an initial draft from my body-worn camera audio. I reviewed the draft, verified every factual claim against the footage, corrected errors, and adopted this account as my sworn statement."

That answer is accurate, complete, and defensible. It does not hide the AI involvement. It does not cede authorship to the machine. It explains the process in terms a judge, prosecutor, and defense attorney can evaluate. Officers should practice that answer in training. Role-play the deposition question three times per officer, with different framings of the question from the trainer. The familiarity reduces the chance that an officer will be caught off guard in a real proceeding and give a less precise answer that creates unnecessary doubt about the report's integrity.

Dispatch: The Call-Handling Curriculum

Telecommunicators operate in the highest-stakes real-time environment in public-safety AI. A misclassified call is not a documentation error. It is a priority error, and a priority error in dispatch can mean the wrong resources arrive at a life-threatening situation or the right resources are delayed. The dispatch curriculum is built around the principle that the telecommunicator owns the classification, the AI assists with documentation, and no AI output ever changes the priority without the telecommunicator's explicit confirmation.

Call Summary Verification

AI-assisted call summarization takes the audio of an incoming call and produces a CAD entry: call type, caller information, location, relevant details, and preliminary priority. The telecommunicator's task is to verify that summary against the call recording before it is entered into the CAD and before units are dispatched. The verification is faster than the patrol verification pass because the call is usually shorter and the claim density is lower, but the consequence of a missed error is immediate and potentially irreversible.

Dispatch training should focus on the specific error types that appear in AI-generated call summaries. The most common: misclassified call type (the AI categorizes a domestic disturbance as a welfare check), incorrect location (a garbled address from a caller in distress is transcribed incorrectly), and inferred details that the caller did not actually state (the model fills in "suspect is on foot" because most calls of this type involve a suspect on foot, when the caller never provided that information).

Training should use recordings from the agency's own CAD system, with any identifying information removed. The trainer plays a call and shows the AI-generated summary side by side. Telecommunicators identify discrepancies. The training coordinator should build a set of recordings that includes the most common error types the agency's tool produces, because a telecommunicator who has seen the specific error types their system tends to make is significantly more likely to catch them on the console than one who has only been told errors can occur.

The Human Owns the Priority

The non-negotiable in dispatch AI is that priority classification is a human decision. The AI can assist with documentation, can suggest a call type based on audio content, can help a busy telecommunicator draft a summary. It does not set the priority. The telecommunicator sets the priority, every time, based on their professional judgment and the information in the call. If the AI suggestion and the telecommunicator's judgment diverge, the telecommunicator's judgment governs.

This needs to be stated explicitly in training and reinforced in policy. Telecommunicators who work in high-volume environments are under constant time pressure. A tool that produces a suggested priority can create subtle automation bias: the tendency to accept a suggested classification because challenging it takes cognitive effort in a moment when cognitive effort is already rationed across multiple open incidents. The training needs to name that bias explicitly and provide a simple protocol: the telecommunicator confirms the priority, every time, before dispatch, regardless of what the AI suggested.

CAD Entry Accuracy and the Incident Record

The CAD entry is not just a dispatch tool. It is the beginning of the incident record. Patrol officers who arrive at a scene use the CAD entry to understand the nature of the call. Supervisors use it to track unit allocation. Records clerks use it as the anchor for the case file. Investigators use it to reconstruct the timeline. An inaccurate CAD entry propagates error through the entire chain of documentation that follows it.

Telecommunicator training should make this propagation explicit. Show a case in which an inaccurate CAD entry affected the patrol report, the case file, and the prosecutor's charging decision. The point is not to create anxiety. It is to demonstrate that verification at the dispatch stage prevents errors at every subsequent stage. An accurate CAD entry is the foundation of an accurate case record, and the telecommunicator is the person who ensures that foundation is sound.

Records: The Redaction and Release Curriculum

Records clerks occupy a position in the AI program that is structurally different from patrol and dispatch. They are not generating content from live incidents. They are processing completed case files, fulfilling public-records requests under open-government statutes, and applying AI-assisted redaction to body-worn camera footage and documentary records. Their failure mode is not a gap-fill in a sworn narrative. It is a redaction failure: a missed face, a visible license plate, an unredacted witness address in a released document.

AI-Assisted Redaction and Its Verification

AI-assisted redaction tools identify faces, license plates, and other sensitive information in video and documents and apply redaction marks. The tools are significantly faster than manual redaction. They are not infallible. A face at an angle, a partially obscured plate, a name in a handwritten note: these are the categories where automated redaction tends to miss and where a missed redaction has legal consequences. Under open-records and public-records statutes, releasing footage with an unredacted face of a minor, a victim, or a confidential informant is not a minor error. It is a privacy violation, a potential lawsuit, and in some cases a danger to the person whose identity was exposed.

The records curriculum should include a verification protocol for AI-assisted redaction that is as rigorous as the patrol officer's footage-grounded verification pass, adapted to the records context. The verification pass for redaction is a frame-by-frame review of the redacted footage, with specific attention to the categories the automated tool tends to miss. Training should use examples from the agency's own redaction tool, including examples of redaction errors that the tool has produced, because records clerks who have seen the specific failure modes of their specific tool are substantially better at catching them than clerks who have been told errors can occur.

Over-Redaction and Under-Redaction: Both Failures

Records training needs to address a failure mode that is less discussed than missed redaction: over-redaction. An AI redaction tool set to high sensitivity may redact faces that are not subject to privacy protection (bystanders in a public space who have not been identified as witnesses, for example) or may redact information that is subject to public disclosure under open-government statutes. Over-redaction that withholds information the public or the defense is legally entitled to receive is not a safe error. It is a different class of failure with its own legal consequences.

Records clerks should be trained on the agency's specific redaction policy: what categories of information require redaction, what categories do not, and what the standard is for borderline cases. The CJIS (Criminal Justice Information Services) Security Policy, which governs the handling of criminal justice information, sets baseline standards for data protection but does not prescribe redaction decisions in individual cases. Those decisions are governed by state open-records statutes, agency policy, and case-specific circumstances. Records clerks need to understand that framework well enough to make or escalate borderline calls, not just to run the automated tool.

The Public-Records Request Workflow

AI assistance in the public-records workflow extends beyond redaction to summarization and tracking. Records departments that process high volumes of requests use AI to categorize incoming requests, draft acknowledgment letters, track deadlines, and summarize case materials for review. Each of those uses has its own verification requirement.

Training should walk through the full request-to-release workflow, identifying at each step where AI assists and where human verification is required. A records clerk who understands the full workflow, not just their specific step in it, is better positioned to catch errors that originate at one step and become problems at another. The clerk who verifies the redaction but does not notice that the AI summarization misidentified the requester's statutory basis for the request may release the right footage for the wrong reason, which can create its own legal complications.

Making Training Stick Across the Agency

Initial training is necessary but not sufficient. The verification habits that an officer, telecommunicator, or records clerk builds in a forty-five minute session will erode under operational pressure unless the agency builds reinforcement structures into the workflow.

The most effective reinforcement structure is supervisor review with a verification focus. Supervisors who review AI-assisted reports should specifically check for the presence of the disclosure note, the completeness of the verification log, and whether any corrections documented in the log were significant enough to warrant a supervisor notification. This is not a punitive review. It is a quality-assurance review that signals to the officer that the verification standard is real and that it is being monitored.

Peer learning is the second reinforcement structure. The agency should designate at least one AI champion per shift per unit: an officer, telecommunicator, or records clerk who has completed the full curriculum, is available to answer questions from colleagues, and can provide peer-level guidance when someone is uncertain about how to handle a specific situation. Champions are the practical resource for the moments when training is not fresh and the standard is unclear. They are also the peer-testimony resource described in the trust-building lesson: the colleague who can say "I was skeptical, here is what I actually found when I used it."

Refresher training should be scheduled at six-month intervals, or immediately following any incident in which an AI-assisted report, CAD entry, or redaction produced a documented error that reached a legal proceeding. The incident-triggered refresher is the most effective because it is grounded in a real case from the agency's own experience. An officer who hears about a gap-fill that a colleague missed in a use-of-force report and that came up in a suppression motion is more attentive to gap-fills in their own drafts than an officer who hears about a hypothetical.

Finally, the training coordinator should maintain a running log of the errors that AI tools produce in the agency's specific operational environment, because those errors are the most relevant training material available. A lesson that uses a generic example of a gap-fill is useful. A lesson that uses an actual gap-fill from the agency's own deployment, with the officer's permission and identifying information removed, is significantly more useful. The training gets better as the deployment matures, if the training coordinator is collecting the data that makes it better.

Key Takeaways

• Role-specific training is the minimum standard. Patrol officers, telecommunicators, and records clerks each operate in a distinct evidentiary environment with distinct AI output types, distinct failure modes, and distinct verification requirements. A single generic session does not meet that standard.
• The four patrol competencies are gap-fill recognition, the footage-grounded verification pass, disclosure language and audit logging, and the deposition answer. All four must be trained and practiced, not just described.
• Dispatch training must name automation bias explicitly and build a protocol that ensures the telecommunicator confirms the priority classification on every call before dispatch, regardless of the AI suggestion.
• Records training must address both under-redaction (the missed face that is a privacy violation) and over-redaction (the withheld information the requestor is legally entitled to receive). Both are failures with legal consequences.
• CJIS Security Policy obligations remain with the agency. The verification log, the disclosure note, and the audit trail are part of meeting those obligations, and records clerks need to understand the framework within which they operate, not just how to run the redaction tool.
• Supervisor review focused on verification quality, not report speed, is the most effective reinforcement structure for sustaining the verification habit under operational pressure.
• Per-shift AI champions provide peer-level guidance and peer testimony that no amount of command-staff messaging can replicate. They are a structural investment in sustained competence.
• Training gets better as the deployment matures if the agency is systematically collecting real error data from its own tools and using that data to update its training scenarios. Generic training is the starting point; agency-specific training is the standard.