The defense attorney waited until the detective had been on the stand for forty minutes before she asked the question. The detective had walked through the timeline, identified the exhibits, and explained the investigation in clean, professional testimony. Then: "Detective, who wrote Exhibit 14?" The detective paused. Exhibit 14 was the incident report narrative. "I did," the detective said. The attorney nodded slowly. "Did you use any software to assist you in drafting the narrative?" Another pause. "I used a drafting tool." "What tool?" "I believe it was a product from the camera vendor." "Did you disclose that to the prosecutor?" The pause this time was long enough for the jury to notice.

The detective had done nothing wrong operationally. The report was accurate. The verification had been thorough. The problem was that the detective had no framework for answering the question, no documentation to point to, and no ready vocabulary for explaining, under oath and in plain language, exactly what the AI tool had done and what the detective had done in review. That gap, between the legitimate use of an AI drafting tool and the ability to account for that use on the stand, is the gap this lesson closes.

Admissibility is the legal determination that a piece of evidence or testimony is proper to be presented to the finder of fact (the judge or jury) in a legal proceeding. Evidence and testimony can be excluded on many grounds: lack of foundation, hearsay, improper authentication, unreliable methodology, and prejudice that outweighs probative value. For a police report drafted with AI assistance, and for the officer who testifies about that report, the admissibility questions cluster around three concerns: was the report accurately authored, was the process reliable, and can the officer establish the foundation for the report's contents as their sworn account?

The Deposition and What It Actually Asks

A deposition is a pretrial procedure in which a witness testifies under oath, outside the courtroom, in response to questions from the attorneys. The testimony is recorded and can be used at trial to contradict the witness if their trial testimony differs, or in some cases as a substitute for live testimony. Officers are frequently deposed in civil cases arising from their enforcement actions, and in serious criminal cases the defense may depose officers as well. Depositions in cases involving AI-assisted reports are adding a new line of questioning that officers need to be able to answer without hesitation.

The core deposition question is not "did you use AI?" That question is a predicate. The real questions are these: "What did the AI produce?", "What did you review?", "What did you change?", "How did you verify the accuracy of what you submitted?", and "Is every factual statement in this report something you personally can vouch for?" An officer who can answer all five of those questions clearly, with supporting documentation in hand, is a strong witness. An officer who fumbles any one of them is a vulnerability for the prosecution and an opportunity for the defense.

The difficulty is that the AI drafting workflow, as it exists today, does not automatically generate the documentation that would make those answers easy. Axon's Draft One, which produces report narratives from body-worn camera (BWC) audio, provides the officer with a draft. Whether the agency's workflow requires preservation of the original draft, a log of the officer's edits, and a notation in the case file indicating the tool was used is a policy question that varies by agency. In the absence of that policy, officers are testifying about their AI-assisted reports from memory, which is exactly the situation the detective above found herself in.

The officer who can walk a jury through each verification step is a stronger witness than the officer who used the same tool and cannot describe the review at all.

The Four Questions Every Officer Must Be Able to Answer

Based on the emerging pattern of cross-examination and deposition questioning around AI-assisted police reports, there are four questions every officer should be prepared to answer before they submit a report drafted with AI assistance.

Question one: "Did you use AI to help draft this report?" The answer to this question should always be yes, if you did, stated clearly and without hesitation or shame. Using AI to assist report drafting is not misconduct. Hiding that you did it, or being unable to account for it, is the problem. An officer who has a clear agency policy authorizing AI use and a documented verification practice has nothing to hide and everything to gain from transparency on this point.

Question two: "What did the AI produce, and what did you change?" The officer needs to be able to describe the AI tool's function at a general level (it drafted an initial narrative from my body-worn camera audio), and ideally to identify specific changes made to the draft. If the original draft has been preserved in the case file, as it should be under a proper retention policy, the officer can point to it. If the original draft has been deleted, the officer is testifying from memory about a document that no longer exists, which is a weak position and one that suggests the agency has a records-management gap.

Question three: "How did you verify the accuracy of the report before you submitted it?" This is the question that determines whether the officer is a credible witness or a vulnerable one. The answer must be specific: "I reviewed the draft narrative against my body-worn camera footage, paragraph by paragraph. I checked the timeline against the computer-aided dispatch (CAD, the system that logs every dispatch timestamp and action) log. I corrected three discrepancies: the model had slightly paraphrased a statement I made verbatim to the suspect at the initial contact, the model had omitted the description of the second witness, and the model had transposed the address numbers. I corrected all three before submission." That answer, given confidently and supported by a verification log, is impenetrable cross-examination territory for the defense.

Question four: "Is everything in this report true and accurate to the best of your knowledge?" The answer must be yes, and it must be supported by the verification work that preceded submission. An officer who says "yes" to this question and means it, because they did the work, is in a fundamentally different position from an officer who says "yes" because they cannot imagine saying no. The sworn certification of a police report is a serious act. It means the officer has personal knowledge of, or has personally verified, every material statement in the document.

Foundation and Authentication for AI-Assisted Reports

Before a document can be admitted into evidence, the party offering it must establish its foundation: the factual predicate showing that the document is what it is claimed to be, that it was made in the regular course of business, and that it is reliable. For police reports offered under the public records exception to the hearsay rule (Federal Rule of Evidence 803(8) and its state equivalents), the foundation requirements include that the report was made by a public official with a duty to make it, that it was made from personal knowledge or from information transmitted by someone with personal knowledge, and that the record was made at or near the time of the event.

AI-assisted reports add a new layer to the foundation analysis. If the defense objects that the report lacks adequate foundation because the officer cannot establish personal knowledge of every statement in the document, the prosecution must be able to show that the officer reviewed the AI-generated draft, verified its contents against the source record, and adopted the final document as their own sworn account. The adoption is the legal hinge. A draft that the officer reviewed and adopted as their sworn statement is the officer's document. A draft that the officer glanced at and submitted without meaningful review is a document the officer cannot fully vouch for, and the foundation for admitting it as a public record may be contestable.

The concept of adoption here is worth dwelling on. When an officer takes an AI-generated draft and submits it as a police report, they are not just formatting a document: they are adopting the statements in it as their own sworn account of what happened. That adoption is a legal and ethical act. It carries the same weight as a handwritten report signed by the officer, because from the court's perspective, it is exactly that. The defense will test whether the adoption was genuine, meaning the officer can substantiate each factual claim, or merely formal, meaning the officer clicked submit on an unreviewed document. The verification record is what distinguishes one from the other.

Chain of Custody for AI-Generated Material

Chain of custody describes the documented and unbroken sequence of possession, handling, and transfer of a piece of evidence from collection through trial. For physical evidence, chain of custody protocols are well-established: the evidence is logged, tagged, stored, and each transfer is signed. For digital evidence including BWC footage, the evidence management platform typically maintains an immutable log.

For AI-generated intermediate documents, chain of custody is less well-developed. Who has access to the AI output? Can it be modified after generation? Is there a system-level log of changes? Is the final submitted report distinguishable from an earlier draft? These questions are not hypothetical. Defense forensic experts in cases involving AI-assisted documentation are beginning to request system logs from AI platforms as part of discovery, and where those logs show that the document was modified after its initial generation in ways the officer cannot account for, the chain-of-custody argument becomes available to the defense.

Officers cannot individually control the architecture of their agency's AI evidence-management system. But they can take three steps that protect chain of custody at the individual level: they can preserve their own copy of the AI-generated draft before editing, they can document their editing process contemporaneously, and they can ensure that the submitted report is clearly identified as the final, officer-adopted version that supersedes the AI draft. These steps are not bureaucratic overhead. They are the professional practice of an officer who understands that every document touching a criminal case is evidence.

Cross-Examination and How It Uses AI Against Officers

Cross-examination is the questioning of a witness by the opposing party, and it is the most adversarial moment in a trial. Defense attorneys are trained and practiced at finding inconsistencies, using prior statements against witnesses, and undermining the credibility of government witnesses through specific and pointed questions. AI-assisted reports give defense attorneys several new angles of attack.

The discrepancy attack. The defense reviews the AI-assisted report against the BWC footage (which is in discovery), finds a detail in the report that differs from what the footage shows, and asks the officer about it on the stand. The officer may not remember the detail independently because the report was drafted from the footage by an AI and reviewed months before trial. The officer's inability to explain how the discrepancy got there, or to say "that was in the AI draft and I missed it in review," is a credibility problem. The prevention is the verification log: if the officer verified every factual claim before submission and the log shows no discrepancy was present at verification, the officer can say so. If the discrepancy entered through a process the officer cannot account for, the cross-examination will be uncomfortable.

The authorship attack. The defense establishes that the report was drafted by an AI, asks whether the officer read the entire draft carefully, and then finds a sentence or phrase in the report that is inconsistent with the officer's characterization of their review. This attack works best against an officer who testified that they "went through the report carefully" but cannot describe the specific review process they used. The prevention is the same: a documented, specific verification process that the officer can articulate in detail.

The reliability attack. The defense calls an expert witness on AI reliability, or simply reads into the record documented instances where AI report-drafting tools have produced errors, and asks whether the officer took steps to independently verify the tool's output against the primary sources. If the officer did, and can describe those steps, this attack fails. If the officer did not, and relies on the AI's general reliability as a defense, they are in a difficult position: the point of the attack is exactly that AI reliability is not a substitute for officer verification.

The disclosure attack. The defense argues that the prosecution failed to disclose the AI tool's involvement, the AI draft, or the officer's review log as required by Brady, Giglio, or procedural rules. This attack is not about the report's accuracy; it is about the integrity of the disclosure process. It works by transforming the trial from a question about what happened on the night in question into a question about whether the prosecution played by the rules. An agency with a documented, consistent AI disclosure policy largely forecloses this attack. An agency without one is vulnerable to it every time a case goes to trial.

Building Testimony-Ready Documentation

The officer who wants to be a strong witness in AI-related cross-examination needs to build that strength before the incident, not during the deposition. The documentation practices described here are not extra work layered on top of the job. They are a shift in where in the workflow the documentation happens, earlier and more specifically, in exchange for confidence at the end when it matters most.

The verification log. Every time an officer reviews an AI-generated draft against the source record (BWC footage, CAD log, witness notes), they should make a brief contemporaneous notation of what was checked and what was found. The notation does not need to be elaborate: a field in the RMS (records management system, the agency's central case-file database), a timestamped note in the case file, or a checked-box verification form attached to the draft. The notation needs to exist and be dated before submission. A verification log made the day before trial is not the same as a verification log made the day of submission.

The draft retention record. The AI-generated draft before officer edits should be retained in the evidence management system and linked to the case file. Many agency platforms that integrate AI report-drafting tools have the technical capability to retain draft versions; whether the agency has turned on that retention setting and built it into policy is the question. Officers should know whether their agency retains AI drafts, and if the answer is no, they should flag it through the appropriate command channel as a policy gap.

The disclosure notation. Every case file that involved AI assistance should include a notation to the assigned prosecutor, delivered before the discovery package is prepared. The notation should identify the tool, the scope of use, and the review process. This is a brief communication, not a formal report, but it creates the paper trail that allows the prosecutor to make an informed disclosure decision and document that the decision was made with full information.

The sworn-statement readiness check. Before an officer submits an AI-assisted report, they should ask themselves the following question: can I testify, under oath, about every factual statement in this document? If the answer to any particular statement is "I believe the AI got that right" rather than "I verified that against the footage and the CAD log," the report is not ready for submission. The AI got many things right. But "I believe the AI got that right" is not the same as "I know this is accurate because I checked it." The standard for a sworn police report is the second one.

Key Takeaways

• Admissibility of an AI-assisted police report turns on whether the officer can establish foundation for the document as their sworn account, based on personal knowledge or verified review, not on whether AI was used. The adoption of the AI draft as a sworn statement is the legal hinge.
• The four deposition questions every officer using AI assistance must be able to answer confidently: Did you use AI? What did it produce and what did you change? How did you verify accuracy before submission? Is every statement in this report true and accurate? Preparation for these questions starts in the verification step, not in the witness box.
• Cross-examination attacks on AI-assisted reports typically take one of four forms: discrepancy attacks (report differs from footage), authorship attacks (officer cannot describe the review), reliability attacks (AI tool's accuracy is challenged), and disclosure attacks (Brady/Giglio non-disclosure). Rigorous documentation and early disclosure close three of the four.
• A deposition (pretrial sworn testimony by a witness) and cross-examination are adversarial proceedings designed to surface inconsistencies. An officer who can point to a contemporaneous verification log is a far stronger witness than one who is reconstructing their review process from memory months after the fact.
• Chain of custody discipline for AI-generated documents means preserving the original draft before editing, documenting the editing process, and clearly designating the final submitted version as the officer-adopted account. Officers cannot control the platform's architecture, but they can build their own contemporaneous record.
• The verification log, the draft retention record, and the disclosure notation to the prosecutor are the three documentation practices that separate testimony-ready AI use from legally exposed AI use. They require no new technology: a field in the RMS and a two-sentence email to the prosecutor are sufficient at minimum.
• Transparency about AI tool use, documented and proactive, is a credibility asset in court. An officer who says clearly and without hesitation "yes, I used a drafting tool, here is what it produced, here is what I verified, and here is what I adopted as my sworn account" is giving the defense nothing to work with. Concealment or confusion about AI use gives the defense exactly the question it needs.