Officer Renata Flores finished her fourth call of the night, a domestic disturbance that left her with a witness statement, a CAD (computer-aided dispatch) entry, forty-two minutes of body-worn camera (BWC) footage, and the quiet knowledge that the report she was about to write would matter in court. Her agency had deployed Axon Draft One three months earlier. She opened the platform, clicked "Generate Draft," and ninety seconds later had a two-page narrative sitting in her records management system (RMS, the agency's central repository for offense reports and case files). It read cleanly. It sounded like her. It was, she knew, not yet hers.

The Draft Is a Starting Point, Not a Finish Line

The single most important concept in this lesson is the one that almost every new user of AI-assisted report writing gets wrong. The AI draft is not a report. It is a starting point, a structured proposal built from audio and pattern-matching, waiting for the one person who was actually there to read it, correct it, and own it. Until the officer reads every line, checks every claim, and affirmatively adopts the document as their sworn account, it is nobody's report. It is a computer's guess.

That framing is not rhetorical. In testing that accompanied Axon Draft One's rollout, officers reported an 82% decrease in report-writing time. That number is real and worth pursuing. But the 82% savings is realized on the drafting end, not the review end. The officer still has to do the review. The review is the job now. Everything else is scaffolding.

What changes when AI drafts the report is where your cognitive effort goes. Before AI, you spent the bulk of the time constructing sentences, finding the right words, struggling to get the sequence right after a chaotic call. After AI, the sentences are constructed for you. The struggle moves upstream: does this sentence describe what actually happened, or does it describe what usually happens in calls like this one? That is a harder question. It requires a different kind of attention. And it is, in the end, a better use of your expertise than formatting prose.

What Adoption Means Legally

Adoption is not scrolling to the bottom and clicking "submit." It is a deliberate act with legal consequences. When you adopt an AI draft as your sworn report, you are representing to your agency, to the prosecutor's office, to the defense, and to any court that reads the document that the narrative reflects your personal knowledge and observation of the incident. The word "sworn" in "sworn report" means exactly that: you are swearing the contents are true.

That obligation does not change because you used a tool to generate the first draft. A carpenter who uses a nail gun is still responsible for the quality of the joint. A doctor who uses a diagnostic algorithm is still responsible for the diagnosis. An officer who uses an AI platform to draft a report is still the author of that report, and "the AI wrote it" is never an acceptable answer in a deposition, a suppression hearing, or a Brady (Brady v. Maryland, 373 U.S. 83) disclosure review.

Brady v. Maryland and its companion case Giglio v. United States (405 U.S. 150) together require that the prosecution disclose all material exculpatory evidence and all evidence that could be used to impeach a witness, including the officer who wrote the report. An AI-drafted narrative that contains a fabricated detail, a softened description of force, or a misquoted statement is a potential Brady problem the moment it enters the file. The defense has the right to probe how the report was written. If the answer is "AI generated it and I didn't fully verify it," the officer's credibility and the case are both at risk.

The AI draft becomes a sworn report at the moment you adopt it. From that moment forward, you are the author of every word in it, and you are accountable for every claim it makes.

The Five-Phase Review Discipline

A good adoption review has structure. Officers who approach it casually, reading through once and looking for typos, are not doing the review the document requires. Officers who follow a deliberate sequence catch things the casual read misses. Here is the sequence that works.

Phase One: Read for Sequence

Before you check any individual fact, read the draft straight through and ask one question: is the sequence of events correct? AI systems that work from audio transcription will generally produce a narrative that follows the audio timeline, but audio is not always the most important timeline. The CAD entry has its own timestamp. The BWC footage has timestamps embedded. The witness's account has a sequence. A suspect's statements have a sequence. Do they agree with each other and with the narrative the AI has produced?

Sequence errors in reports are a defense attorney's leverage. If your narrative says the suspect was placed in handcuffs before he made a statement, but the BWC footage clearly shows the statement was made before the handcuffs went on, you have created an inconsistency that will be used to question your credibility on every other claim in the report. The AI did not intend to create that inconsistency. It assembled the narrative from audio patterns and lost the precise sequence in the assembly. Your job is to catch it before it becomes a problem.

Phase Two: Verify Every Fact Claim

Go through the draft sentence by sentence and identify every claim that is a fact: a name, a location, a time, a description of an object, a description of an action, a statement attributed to any person. For each fact claim, ask: where does this come from? The BWC footage? The CAD entry? Your personal observation? The witness's statement?

If you can answer that question, you have a verified fact. If you cannot, you have a gap-fill. The gap-fill is the most dangerous failure mode in AI-assisted reporting, and it is discussed in detail in the lesson on common failure modes. For now: any factual claim you cannot point to a specific source for needs to be either verified against the record or removed from the narrative.

This phase takes time. It is supposed to. The time you spend here is the time you are not spending in a deposition explaining why a fact in your report has no evidentiary foundation.

Phase Three: Check Every Attributed Statement

Any time the draft puts words in quotation marks or attributes a statement to a specific person, go back to the audio. Find the moment in the footage where that statement was made. Confirm the words in the draft match the words in the recording. This is not a paraphrase check. This is a verbatim check.

Language models that draft from audio sometimes smooth out what a person actually said. A suspect who said "I didn't do nothing" may appear in the draft as saying "I didn't do anything." Those are not the same statement. The tense, the grammar, the specific word choice can all matter in a cross-examination. If you cannot verify a quoted statement against the audio, either replace it with a paraphrase that accurately captures the substance, or flag it with a note in your field notes and correct the draft.

Phase Four: Verify Descriptive Claims

Descriptions of physical evidence, injury, demeanor, and environment are another area where AI drafts can drift from the footage. A model working from audio cannot see what the camera saw. It may produce descriptions that are generically plausible for the call type but do not match what was actually present. Did the draft describe the scene as "in disarray" when your footage shows a reasonably tidy apartment? Did it describe an injury as "minor" when the footage shows something more significant? Did it describe the lighting conditions accurately?

These details matter not just for accuracy but because they are the things a defense attorney will pull up on a courtroom screen alongside the BWC footage to ask you whether your report and your camera tell the same story.

Phase Five: Confirm Completeness

The final phase is looking for what is absent. AI drafts tend to include what was said on camera and can miss significant things that happened outside the frame, things you observed with your own eyes, or things that happened before the camera was activated. Check your field notes. Check the CAD entry. Is there anything in the incident that the draft does not address? Omissions in a police report can be as damaging as errors, particularly if a defense attorney later argues that the officer failed to document an exculpatory observation.

The Adoption Statement and Why It Matters

Many agencies that have deployed AI-assisted report writing are adding an adoption statement to the report workflow. The statement is short, typically a sentence or two, and it records the officer's affirmative act of review and adoption. A typical formulation looks something like this: "I have reviewed the AI-generated draft of this report against the body-worn camera footage, the CAD entry, and my personal notes. I have verified, corrected, and adopted the narrative as my sworn account." Some agencies add the specific AI tool used and the timestamp of the review.

That statement does several things at once. It creates an audit trail. It documents that the officer performed the verification pass rather than simply accepting the draft. It gives the prosecutor a foundation for explaining to the defense, if asked, how the report was generated and reviewed. And it gives the officer, if deposed, a specific, documented answer to the question "how did you write this report?"

The King County, Washington prosecutor's office drew a sharp governance line in 2025, barring AI-written police reports from their charging process absent specific documentation of the review process. Other prosecutors have followed or are actively considering similar positions. An adoption statement is not just good practice for the officer's protection. It is rapidly becoming a prerequisite for the report to survive the charging and discovery process intact.

If your agency does not yet have a formal adoption statement in the workflow, use your field notes to document the review anyway. Write the date, the time, and the specific steps you took. That documentation is yours regardless of what the agency's form includes.

Corrections and the Author's Voice

When you find something wrong in the draft, fix it. That sounds obvious, but there are two common failure patterns here that are worth naming.

The first is hesitation to change the AI's language because it sounds more polished or professional. Officers sometimes feel that their own words are less good than the AI's words, and they leave the AI language in place even when it does not accurately reflect what happened. This is exactly backwards. A police report is not a writing contest. A report that accurately describes what occurred in your words, including the imperfect and sometimes awkward words of direct observation, is a stronger document than a polished narrative that drifts from the truth. Your voice, your specific observations, your verbs matter. The AI's prose is a scaffold; use the parts that fit and replace the rest.

The second pattern is making corrections so minimal that the errors survive. If the AI wrote that a statement was made at a specific location and you know the statement was made elsewhere, change the location. If the AI attributed a statement to the wrong person, correct the attribution. Do not soften the correction or work around it with qualifying language. Change the claim to the correct claim, and note in your field notes what you changed and why. That note is your protection and your record.

When to Restart Rather Than Revise

There are cases where the draft is so substantially wrong that revising it is harder than starting over. If the AI has misidentified the core sequence of events, or attributed key statements to the wrong persons, or introduced facts about a generic incident type that have no relationship to what your BWC footage actually shows, you may be better served starting the narrative from scratch. Extensive revision of a deeply flawed draft can leave seams in the document: the corrected sections do not always read consistently with the original sections, and a defense attorney reading carefully may spot the inconsistency.

The test is whether you can get from the draft to an accurate sworn account faster by revising than by rewriting. If the answer is no, rewrite. The AI draft served its purpose as a rough frame; it does not have to be the final structure. Officers who feel they must justify using AI by handing in every draft the tool produces are missing the point. The tool is there to save time. If the draft costs more time to fix than to replace, replace it and note in your workflow log what you did.

Handling the First Deposition Question

You will, at some point, sit in a deposition chair and be asked: "Officer, did you write this report, or did a computer write it?" How you answer that question depends entirely on what you actually did when you reviewed and adopted the draft.

If you performed the five-phase review, made corrections, and adopted the document as your sworn account, the answer is direct: "I am the author of this report. I reviewed an AI-generated draft against the body-worn camera footage and the CAD entry. I verified each fact claim, corrected several items, and adopted the narrative as my sworn account. The report reflects my observations." That answer is honest, documented, and defensible. The defense attorney can probe your review process. They cannot shake your authorship, because the authorship is real.

If you accepted the draft without verification, the deposition goes differently. "I reviewed it" becomes hard to defend when the attorney asks you to explain specifically what you checked. An answer that amounts to "I read through it and it looked right" is not a review that meets the standard a sworn report requires. The attorney who can show that the report contains a fact you could not have personally observed, and that you did not catch and correct it, has an opening to argue that you do not actually know what is in the report you swore to.

The Electronic Frontier Foundation (EFF) has raised a direct transparency concern about AI-drafted police reports: the public, and specifically criminal defendants, have an interest in knowing when government documents that can deprive them of liberty were generated by an automated system rather than by direct human observation and writing. That concern is legitimate and it deserves a direct response, not a dismissal. The response is the review: if you can demonstrate, with documentation, that a trained officer reviewed every claim in the AI draft against the primary evidence, the transparency objection is substantially answered. If you cannot demonstrate that, the objection stands.

Key Takeaways

• An AI draft is a starting point, not a report. It becomes a sworn report only at the moment the officer reviews, corrects, and affirmatively adopts it as their personal account of the incident.
• Adoption means the officer can account for every factual claim in the narrative: where it came from, whether it is accurate, and whether it appears in the primary record (BWC footage, CAD entry, field notes, or witness statements).
• The five-phase review covers sequence, fact claims, attributed statements, descriptive claims, and completeness. Each phase catches a different category of AI error. Doing only the first phase is not doing the review.
• Brady v. Maryland and Giglio v. United States frame AI-assisted reporting as a constitutional matter: an AI-introduced error that reaches the sworn report is a potential Brady violation the moment the report enters the case file.
• An adoption statement, whether agency-mandated or officer-created, documents the review process and gives the officer a specific, documented answer for the deposition question about how the report was written.
• The King County prosecutor's bar on AI-written reports underscores that disclosure and documentation of the AI-assistance and review process is not optional in jurisdictions where prosecutors are paying attention to how reports are generated.
• When an AI draft is substantially inaccurate, rewriting from scratch is better than extensive revision. The draft is a tool, not an obligation. The obligation is the accurate sworn account.
• Axon Draft One testing showed an 82% reduction in report-writing time. That saving is real but it comes with a responsibility: the time saved on drafting must be partially reinvested in the systematic verification pass the AI cannot perform for you.