Deputy prosecutor Marcus Hale read the police report twice before he picked up the phone. It was clean. The narrative flowed, the elements of the charge were all present, the timeline was tidy. It was, in fact, a little too tidy for a 2 a.m. disorderly-conduct call that the responding officer had described over the radio as "a complete circus." Hale called the officer, Sergeant Dana Whitfield, and asked a question that, eighteen months earlier, no prosecutor in his county had ever asked: "Did you use Draft One on this one?" There was a pause. "Yeah," Whitfield said. "We all do now." Hale wrote a note in the file: "AI-assisted. No disclosure in report. Verify before charging." Then he opened the email from the elected prosecutor's office, the one every line attorney had received that morning, the one that started with the words "Effective immediately, our office will not accept police reports drafted by artificial intelligence."

Why Disclosure Is the Whole Game

Everything you have learned in this chapter about AI-assisted report writing, the prompting, the footage-grounded verification pass, the act of adopting the draft as your sworn account, comes down to a single test that happens long after you hit submit. A prosecutor, a defense attorney, and eventually a judge will look at your report and ask, in one form or another, how it was made. The honest answer is that an AI tool drafted it from your body-worn camera (BWC, the camera clipped to your chest that records audio and video throughout a contact) audio, and you reviewed and corrected it. The question that decides whether your case survives is not whether you used AI. It is whether you said so.

This lesson is about disclosure: documenting that AI assisted, recording how you reviewed it, and disclosing that fact according to your agency's policy and your prosecutor's policy. It is the fourth non-negotiable of this entire program, and it is the one officers are most tempted to skip, because it feels like volunteering a weakness. It is the opposite. A documented, disclosed, well-reviewed AI-assisted report is a stronger document than a silent one, and the officers who understand why are the ones who will still be testifying confidently five years from now.

The instinct to hide the tool is human and wrong. When you bury the fact that AI drafted your report, you convert a routine workflow detail into a concealment. And concealment, once discovered, contaminates everything. The defense attorney who learns in deposition that you used AI and never said so does not think "minor omission." They think "what else did this officer not tell us?" The tool was never the problem. The silence is the problem.

You are not disclosing a mistake. You are disclosing a method. Methods that are disclosed and documented survive cross-examination. Methods that are hidden do not.

The King County Line

In 2025, the King County, Washington prosecuting attorney's office drew a governance line that every officer in the country should know about by name. The office announced it would not accept police reports written using AI, citing concerns about accuracy, the reliability of the technology, and the difficulty of testing an automated narrative in court. It was not a quiet internal memo. It was a public, on-the-record position from one of the largest prosecutor's offices in the Pacific Northwest, and it landed exactly as agencies were rushing to deploy Axon Draft One, the tool that drafts police report narratives directly from BWC audio.

It is tempting to read the King County position as anti-technology obstruction, and to dismiss it. Do not. The program's fifth non-negotiable is that you must be evenhanded by design, and an officer who cannot articulate why a serious prosecutor would refuse AI-written reports cannot responsibly use the tool. So articulate it. The prosecutor's worry is concrete: a police report is evidence, it is disclosed to the defense, it is tested at trial, and if a machine generated language that the officer did not fully verify, then the prosecutor is putting an untested document in front of a jury and an officer's sworn name on top of it. That is a real risk. The King County office decided the cleanest way to manage it was a flat bar.

Here is the part most officers miss. The King County position is not primarily an objection to AI. It is an objection to undisclosed, unverified AI. The cure for that objection is not to hide the tool more cleverly. The cure is to do the opposite of what the office is worried about: disclose that AI assisted, document exactly how a trained human verified every claim against the record, and hand the prosecutor a report they can defend. This lesson teaches you to meet that objection head-on, in the open, with paper.

What Disclosure Actually Means

Disclosure is a word with several layers in criminal justice, and conflating them is how officers get into trouble. Let us separate them carefully, because each layer has a different audience and a different rule.

Three Layers of Disclosure

Layer one is internal documentation. This is the record you create showing that an AI tool drafted the report and that you reviewed it. It lives in your field notes, in the report metadata, and ideally in an adoption statement attached to the report itself. Its audience is your agency, your supervisor, and your future self in a deposition chair. It exists whether or not anyone ever asks for it.

Layer two is constitutional disclosure. This is the obligation that flows from Brady v. Maryland (373 U.S. 83), which requires the prosecution to disclose to the defense all material evidence favorable to the accused, and Giglio v. United States (405 U.S. 150), which extends that duty to evidence that could be used to impeach a prosecution witness, including the officer who wrote the report. If the way your report was generated bears on its reliability, and a defense attorney could use it to challenge your credibility, then the method of generation can become material to the defense. The prosecutor, not you, decides what gets disclosed under Brady and Giglio, but the prosecutor can only disclose what they know. They can only know what you tell them.

Layer three is public-records disclosure. Under open-records and public-records statutes, members of the public, journalists, and oversight bodies can request police reports. The Electronic Frontier Foundation (EFF, a digital-rights organization that has scrutinized police technology for years) has raised a direct transparency concern here: the public has a legitimate interest in knowing when government documents that can deprive a person of liberty were produced with the help of an automated system. Whether AI assistance appears in a released report depends on agency policy and statute, but the concern is real and the trend is toward more transparency, not less.

These three layers are connected by a simple chain. If you do layer one well, the internal documentation, then layer two becomes possible, because the prosecutor has something to disclose. And layer three becomes defensible, because the agency can answer a public-records request or a journalist's question with a true, documented account instead of a denial that later collapses. Skip layer one, and the other two layers turn into liabilities you did not see coming.

Internal documentation is the foundation. Constitutional disclosure and public-records disclosure are both built on whether you documented the AI's role in the first place.

The Adoption Statement as Disclosure

The single most practical disclosure tool you have is the adoption statement, and it does double duty. You met it in the lesson on reviewing and adopting an AI draft, where it served to document that you performed the review. It is also, in plain terms, your disclosure that AI assisted. The same sentence accomplishes both jobs.

A complete adoption-and-disclosure statement records four things: that an AI tool was used, which tool, what you verified it against, and that you adopt the result as your sworn account. A typical formulation reads like this: "This report was drafted with the assistance of Axon Draft One from body-worn camera audio. I reviewed the draft against the BWC footage, the computer-aided dispatch (CAD) entry, and my field notes; I verified, corrected, and adopted the narrative as my sworn account on [date and time]." That is roughly fifty words. It takes thirty seconds. It is the difference between a deposition you control and a deposition that controls you.

Notice what the statement does and does not claim. It does not say "the AI got it right." It does not vouch for the tool. It says a human used a tool, a human checked the work against the primary record, and a human takes responsibility for the result. That is the only honest claim available, and it is exactly the claim that answers the prosecutor's concern and the EFF's concern at the same time.

Naming the Tool Specifically

Be specific about which tool drafted the report. "AI-assisted" alone is weaker than "drafted with Axon Draft One." The specificity matters for two reasons. First, different tools have different known behaviors, and the prosecutor or the defense may need to know which system produced the language to evaluate it. Second, vague disclosure invites the cross-examination question "what exactly do you mean by AI?" and you do not want to be improvising an answer to that under oath. Name the product, name the version if you can, and move on. Precision is armor.

Recording What You Changed

Strong disclosure goes one step further than "I reviewed it." Where practical, note what you changed. If the draft attributed a statement to the wrong person and you corrected it, a one-line note in your field notes ("draft attributed the threat to the male party; footage at 04:11 shows the female party made it; corrected") is worth its weight in gold. It demonstrates that your review was real and substantive, not a rubber stamp. It is the concrete evidence that distinguishes you from the officer who clicked submit. And if the defense ever argues that you did not actually review the draft, your change log is the document that ends the argument.

You do not need to log every comma. Log the substantive corrections: a fixed sequence, a corrected attribution, a removed gap-fill, a quote you replaced with a paraphrase because you could not verify the exact words. These are the changes that prove the human did the work the machine could not.

Disclosing to the Prosecutor

The prosecutor is your most important disclosure audience, because the prosecutor stands between your report and the courtroom, and because Brady and Giglio obligations run through that office. The rule here is simple and absolute: tell the prosecutor's office the truth about how the report was made, early, and let them make the disclosure calls that are theirs to make.

This is uncomfortable for many officers because it feels like handing the defense ammunition. It is not. The prosecutor cannot comply with Brady and Giglio if they do not know the facts, and a prosecutor blindsided in trial by an AI-assistance fact they did not know about is a prosecutor who will never trust your reports again, and who may have to dismiss a case or face a mistrial. Worse, an undisclosed fact that surfaces late can be characterized by the defense as a Brady violation, which is a constitutional problem that reaches far beyond the single report. You protect the case by telling the prosecutor what they need to know to do their job.

Meeting a Disclosure Demand

Increasingly, prosecutors are not waiting to be told. Following the King County line, some offices are sending out standing demands: any case involving an AI-assisted report must flag it at intake, and the file must include documentation of the review. When your prosecutor's office issues such a demand, treat it as the floor, not the ceiling. Flag the report. Attach the adoption statement. Attach the change log if you have one. If the office wants a specific form, fill out the specific form. The officers who comply cleanly and completely are the officers whose cases get charged; the officers who treat the demand as bureaucratic friction are the officers whose cases get sent back or declined.

If your prosecutor's office has no policy yet, do not take the silence as permission to stay quiet. Disclose anyway. A short note to the deputy handling the case, or a standing practice of including the adoption statement in every AI-assisted report, costs you nothing and protects you completely. The absence of a demand is not the absence of a Brady obligation. The constitutional duty exists whether or not anyone has written a local rule about it.

When the prosecutor has a policy, follow it to the letter. When the prosecutor has no policy, disclose anyway. The Brady duty does not wait for a local rule.

When the Prosecutor Bars AI Reports Entirely

What if your prosecutor takes the King County position and refuses AI-written reports outright? Then you have a clear answer: in cases handled by that office, you do not submit AI-drafted reports, or you submit them only in the form the office will accept. This is not a battle for the patrol officer to fight. The prosecutor sets the terms on which they will take your cases to court, and an officer who ignores that is creating cases that cannot be prosecuted. Raise the operational reality up your chain of command, where command staff and the prosecutor can work out a policy. At your level, the rule is to know your prosecutor's position and to honor it, because a report the prosecutor will not accept is a report that helps no one.

Here is the strategic point worth carrying to command staff, though. The King County bar targets unverified, undisclosed AI use. The most likely path to a workable policy is not arguing that AI reports are fine. It is showing the prosecutor a disclosure-and-verification regime so rigorous that the underlying concern evaporates: every AI-assisted report carries an adoption statement, a documented verification pass against the footage, and a change log. Many of the prosecutors watching this space are open to AI assistance that comes wrapped in exactly that kind of accountability. The agencies that get to keep their 82% report-time reduction will be the ones that bring receipts.

The Deposition and the Witness Stand

Sooner or later, an attorney will ask you under oath how your report was written. The whole point of disclosure by design is that this moment should be boring. You should have nothing to reveal, because you already revealed it, and nothing to defend, because the defense is in the file.

Picture the two versions of this moment. In the first, the defense attorney asks, "Officer, did a computer write this report?" and you answer, "An AI tool drafted it from my body-worn camera audio. I reviewed the draft against the footage, the CAD entry, and my notes; I corrected the errors I found; and I adopted it as my sworn account. That's all documented in the file, including the adoption statement and my change log." The attorney has nowhere to go. Every fact they hoped to spring on the jury is already on the table, framed by you, in your words. The disclosure has defused the ambush.

In the second version, the attorney asks the same question, and you hesitate. Maybe you say "I wrote it" because you are afraid of where the truth leads. Now the attorney produces the agency's Draft One licensing contract, or a metadata field, or a colleague's testimony that the whole squad uses the tool, and asks why you did not mention it. The case is no longer about the disorderly-conduct charge. It is about whether the officer told the truth about how the evidence was made. You have handed the defense a credibility issue, which is precisely the Giglio impeachment material that should never have existed. The lie, not the tool, is the wound.

The Exact Words That Work

Practice the disclosure answer until it is automatic, the way you practice reading Miranda. A reliable structure is: name the tool, name what you verified against, name that you corrected and adopted, and point to the documentation. "Axon Draft One produced the initial draft from my body-worn camera audio. I checked it against the BWC footage, the CAD entry, and my field notes. I corrected several items, including [the specific correction if you have one], and I adopted the narrative as my sworn account. My adoption statement and change log are in the file." Calm, specific, complete. There is no follow-up question that rattles an officer who can say that and mean it.

Notice that this answer is only available to the officer who actually did the work and documented it. You cannot fake your way into it on the stand. That is the deepest reason disclosure by design is not a paperwork chore: the documentation you create at the moment of adoption is the same documentation that protects you years later in a chair you cannot yet imagine. Present-you is writing the script that future-you will read under oath.

Building Disclosure Into the Workflow

Disclosure that depends on remembering to disclose will fail, because under the pressure of a shift, with 30 to 40% of your time already eaten by paperwork, you will forget. The answer is to make disclosure a default that happens by itself, not a discretionary step you have to choose each time. Build it into the workflow so that the disclosed report is the only kind of report you can produce.

What Good Defaults Look Like

• A required adoption statement field. The RMS (records management system, the agency's central repository for offense reports and case files) report form should not let you submit an AI-assisted report without the adoption-and-disclosure statement completed. If the field is mandatory, you cannot forget it.
• Tool metadata captured automatically. The drafting platform should stamp the report with the tool name and version and the timestamp of generation, so the disclosure is corroborated by the system, not just your memory.
• A change-log habit. Make it routine to note substantive corrections in your field notes during the verification pass, while the call is fresh, not reconstructed months later.
• An intake flag for the prosecutor. Where the prosecutor's office wants AI-assisted reports flagged at intake, the case-submission process should carry that flag automatically so it never depends on an individual officer remembering.

These defaults convert disclosure from an act of discipline into a property of the system. The agency that wires disclosure into the RMS and the intake process is the agency whose reports survive discovery as a matter of routine. The agency that leaves disclosure to officer memory is the agency that will eventually have a case blow up over an undisclosed AI-assisted report, and will then scramble to retrofit exactly these defaults after the damage is done.

Who Owns the Policy

The patrol officer or telecommunicator does not write the disclosure policy, but you live inside it, and you should know who does. Command staff, in coordination with the prosecutor's office and often the agency's legal counsel, sets the disclosure standard. The CJIS (Criminal Justice Information Services Security Policy, the federal rules that govern how criminal-justice data is handled) frame the data-handling side, and remember that those obligations stay with your agency, not the vendor, no matter what the Draft One contract promises. Your job is to know your agency's policy and your prosecutor's policy, to follow whichever is stricter, and to disclose even when neither has caught up to the technology yet. When the policies conflict or are silent, you ask up the chain rather than guessing, because a guess about a constitutional disclosure duty is not a guess you want on the record.

The Cost of Getting It Wrong

It helps to be concrete about what failure looks like, because the abstraction "disclose your AI use" does not convey the stakes the way a consequence does. Walk through the chain of events when an officer hides the tool.

The officer submits an AI-assisted report with no disclosure. It contains a gap-fill detail, a fact the model invented to fill a hole the footage did not show, because that is the most common AI failure mode and the officer's verification pass was thin. The case proceeds to charging. The prosecutor, unaware of the AI assistance, builds a case on the report. In discovery, the defense reviews the BWC footage frame by frame and finds the detail that is in the report but not in the footage. The defense then learns, through deposition or through the agency's Draft One contract, that AI drafted the report and the officer never said so. Now there are three separate problems stacked on top of each other: a fabricated fact (a potential Brady violation), an officer who did not verify it (a competence and credibility problem), and an undisclosed method (a transparency and Giglio impeachment problem). The defense files a motion to suppress the report, moves to impeach the officer, and the prosecutor faces the choice of defending a contaminated case or dismissing it.

Every one of those three problems was avoidable, and disclosure addresses all three at the root. The verification pass catches the gap-fill. The documentation proves the verification happened. And the disclosure removes the concealment that turns a workflow detail into a credibility weapon. The officer who discloses by design does not have a perfect report, because no report is perfect, but they have a defensible one, and a defensible report is the only kind worth swearing to.

Contrast that with the cost of doing it right, which is almost nothing: thirty seconds for an adoption statement, a few lines of change log, a flag at intake. The asymmetry is staggering. A trivial, repeatable habit on the front end prevents a case-ending catastrophe on the back end. That asymmetry is the entire argument for disclosure by design.

Key Takeaways

• Disclosure is the program's fourth non-negotiable: document that AI assisted, record how you reviewed it, and disclose per agency and prosecutor policy. The question that decides your case is not whether you used AI but whether you said so.
• The King County, Washington prosecutor's bar on AI-written reports is fundamentally an objection to undisclosed, unverified AI. You meet it head-on by disclosing and documenting a rigorous verification pass, not by hiding the tool more cleverly.
• Disclosure has three layers: internal documentation (your field notes and adoption statement), constitutional disclosure under Brady v. Maryland and Giglio v. United States (the prosecutor decides what to disclose but can only disclose what you tell them), and public-records disclosure (where EFF's transparency concern lives).
• The adoption statement does double duty: it documents your review and it discloses the AI's role. A complete one names the tool, names what you verified against, and affirms you adopted the result as your sworn account.
• Name the tool specifically (Axon Draft One, not just "AI") and keep a change log of substantive corrections. Specificity is armor and the change log proves the human did the work the machine could not.
• Tell the prosecutor the truth about how the report was made, early. The prosecutor cannot meet Brady and Giglio obligations if they do not know the facts, and a prosecutor blindsided at trial is a case lost.
• Practice the deposition answer until it is automatic: name the tool, name what you verified against, name that you corrected and adopted, point to the documentation. That answer is only available to the officer who actually did the work.
• Build disclosure into the workflow as a default (a mandatory adoption-statement field, automatic tool metadata, a change-log habit, an intake flag) so it never depends on memory under the pressure of a shift.