The prosecutor's call came on a Thursday morning, fourteen months after the arrest and nine days before trial. The detective who answered had been promoted twice since writing the report. The report in question had been drafted with AI assistance from body-worn camera (BWC, the recording device on the officer's uniform) audio, submitted, and never flagged. The prosecutor had a problem: a defense investigator had found a detail in the AI-drafted narrative that did not appear anywhere in the footage, the CAD (computer-aided dispatch) entry, or the field notes. The detail was not dramatic. It described the position of an item in a scene in a way that the footage clearly contradicted. It had survived fourteen months in the case file, two prosecutor reviews, and an uncontested preliminary hearing. Now it was nine days from trial, and the question was not whether the error existed. It was what the agency was going to do about it.

What Makes an AI-Related Case Problem Different

Every case has errors. Reports contain mistakes: a transposed address, a misspelled name, a time noted as the wrong hour. Those errors are corrected through normal supplement processes, and they rarely rise to the level of a case problem. An AI-related case problem is different in three specific ways that shape how the response must work.

First, the source of the error may not be traceable. When an officer writes an error by hand, there is a clear authorship chain: the officer wrote it, the officer can explain why, and the correction process is straightforward. When an AI-assisted report contains an error that was not caught in verification, the question of how the error entered the narrative is genuinely uncertain. The AI model may have generated it as a gap-fill, pulling from patterns in its training data to complete an audio gap in the footage. The officer may have inadvertently accepted it during a fatigued review. Or the footage may not have been reviewed at all. The source matters because it determines whether the problem is a one-case error or a systemic pattern that affects every AI-assisted report the agency has ever produced.

Second, the disclosure obligation triggers immediately. Brady v. Maryland (the 1963 Supreme Court case establishing that prosecutors must disclose exculpatory evidence to the defense) and Giglio v. United States (the 1972 case extending that obligation to impeachment evidence, including evidence that could undermine an officer's credibility) do not have a grace period for AI errors. The moment an agency becomes aware that an AI-assisted report contains a material error that could affect the case, the disclosure obligation is active. The question of how the error happened, who is responsible, and what the agency's internal review process will find are all secondary to the immediate obligation to notify the prosecuting office and, through that office, the defense.

Third, the scope of the problem is immediately uncertain. If the AI tool that drafted this report was generating errors through a consistent mechanism, every other report it drafted may have similar problems. A case problem that starts with one error in one report becomes a potential case-review problem across every matter in which the tool was used. That uncertainty is not speculative. It is the first question a reasonable prosecutor, a defense attorney, and an oversight body will ask: "Is this the only one?" The incident response must be designed to answer that question honestly, not to manage the answer.

The first question after an AI case problem surfaces is not "how do we fix this case." It is "how many other cases need the same review, and who do we tell first."

The First Four Hours: Contain the Damage

The immediate response to an identified AI case problem follows a sequence that every agency should have documented before an incident, not assembled in the hours after one. The detective in the opening story spent three of the first four hours trying to find someone with authority to make a decision. That is a governance failure, not a response failure. Incident response requires pre-assigned authority and a pre-documented notification chain.

Step One: Notify the Governance Board and Legal Counsel

The governance board designated in the agency's AI governance charter is the first notification destination, not the last. In practice, the charter should identify a single point of contact who can be reached outside business hours, because AI case problems surface at inconvenient times. That contact person, typically legal counsel or the designated AI program lead, initiates the response and convenes the relevant parties. The notification must be immediate: the disclosure obligation clock starts running at the moment of knowledge, not at the moment the board convenes.

Legal counsel's first task is to assess the disclosure obligation. Does this error meet the Brady threshold as potentially exculpatory? Does it affect the officer's credibility in a way that triggers Giglio? The assessment should be documented in writing from the first moment. If the error appears in a pending case, the prosecutor must be notified within a timeframe that allows them to fulfill their own disclosure obligations to the defense before any proceeding where the error could be outcome-determinative. "We just found out" is not a timing defense if the agency's internal response delayed the external notification.

Step Two: Preserve Everything

Before any review, any interview, any supplemental report, or any correction is made, preserve the original state of every relevant record. This means: the original AI-generated draft before any human editing, the submitted report as filed, the officer's verification notes if any exist, the BWC footage used to generate the draft, the CAD entry, and any intermediate versions of the document if the system maintains version history. Preservation is the step that gets skipped in the urgency of fixing the problem, and it is the step that becomes critical in every subsequent proceeding.

The CJIS (Criminal Justice Information Services) Security Policy obligations the agency bears extend to the preservation of this record. The records exist on the vendor's platform, on the agency's RMS (records management system, the platform where reports are finalized and stored), and potentially in the prosecuting office's discovery files. The preservation must reach all three locations. Contacting the vendor immediately to ensure no automated purge or overwrite occurs on the AI platform side is a technical step that legal counsel and the technical board member should execute within the first hour of knowledge.

Step Three: Document What Is Known and What Is Not

The incident record begins at first knowledge. Every communication, every decision, every finding, and every uncertainty must be documented from the start. This documentation will be the foundation for the board's response to the prosecuting office, the defense, the court, and any subsequent oversight review. It will also be the evidence that the agency's response was reasonable, timely, and transparent, or the evidence that it was not.

The documentation should be structured around three questions: What error exists in the record? What cases does it appear in or may it appear in? Who has been notified and when? Those three questions drive the immediate response and the scope review that follows. An agency that can answer all three within the first 24 hours with documented evidence is in a fundamentally different position than one that can only answer the first.

Scope Review: Finding the Edges of the Problem

The scope review is the most consequential part of the incident response and the part most likely to be underestimated. Its purpose is to determine whether the identified error is an isolated occurrence or a pattern. The difference between those two answers has case, budgetary, legal, and political consequences that no governance board can manage without the facts the scope review produces.

Identify the Population of Affected Reports

The first step of the scope review is to identify every report generated with the same AI tool, in the same call-type category, during the period the tool was in use. This is not a small number in most active deployments. An agency using AI-assisted report writing for 18 months across a high-volume patrol division may have thousands of AI-touched reports in its files. The scope review must establish which of those reports are in active cases, which are in closed cases that resulted in convictions or pleas, and which are in declined or dismissed matters that may become relevant if the error pattern is systemic.

The RMS is the primary data source for this inventory. If the RMS does not tag AI-assisted reports separately from human-written reports, the scope review starts with a fundamental documentation gap that should have been closed at deployment. A governance board that authorized AI-assisted report writing without requiring the RMS to tag AI-generated drafts has made the scope review harder than it needed to be, and harder than any responsible governance framework would have permitted.

The Error Pattern Analysis

Once the population of AI-assisted reports is identified, the error pattern analysis asks: does the identified error reflect a specific failure mode of the AI tool, and if so, which reports are most likely to share the same failure mode? This is technical work that typically requires cooperation with the vendor, because understanding whether the error is a gap-fill from a specific audio condition, a use-of-force boilerplate insertion, or a scene-description inference without audio support requires knowing how the model processes the specific type of footage involved.

The vendor cooperation question is sensitive. The agency needs technical information about the model's behavior to conduct an honest scope review. The vendor's instinct may be to limit what it shares, particularly if the error pattern reflects a systemic model failure rather than a one-time anomaly. The governance board should have anticipated this scenario in the AI use contract, requiring the vendor to cooperate with agency-initiated accuracy reviews as a contract term. An agency that did not include that term is negotiating from a weaker position at exactly the wrong moment.

The error pattern analysis should produce a risk stratification of the AI-assisted report population: reports with high likelihood of containing a similar error, reports with moderate likelihood, and reports that are probably unaffected because the error mechanism does not apply to their call type or footage conditions. That stratification drives the prioritization of the case-by-case review.

Case-by-Case Review Prioritization

The case-by-case review cannot happen simultaneously across all AI-assisted reports in a large agency's files. It must be prioritized. The prioritization framework should rank cases in this order: first, active cases approaching trial where the error could affect the proceeding; second, cases with recent convictions or pleas where the error could support an appeal or post-conviction motion; third, active cases not approaching trial; fourth, closed cases with older convictions; fifth, declined and dismissed matters. Within each category, use-of-force cases and cases with contested factual records should be reviewed before routine matters.

Each case-by-case review should follow the same protocol: retrieve the AI-generated draft (preserved in step two), retrieve the BWC footage, run a footage-grounded comparison on the specific error pattern identified, document what the review found, and assign a disposition: error confirmed, no error found, or inconclusive (footage insufficient to resolve). The disposition drives the next step: cases with confirmed errors proceed to the notification and correction process; cases with no error found are closed in the review record; inconclusive cases require a human investigator to make a judgment call and document it.

Notification, Disclosure, and the Prosecuting Office

The notification sequence is where the incident response becomes a legal and institutional process rather than an internal review. The governing framework is Brady and Giglio, but the practical mechanism is the relationship the agency has (or does not have) with the prosecuting office before the incident.

An agency that built its prosecuting-office relationship before an incident has a direct line to the prosecutor's Brady compliance officer, a shared understanding of what disclosure the prosecutor needs and in what format, and a working relationship that means the prosecutor receives the notification as information rather than as a surprise with an adversarial edge. An agency that did not build that relationship is now establishing it under the worst possible conditions, and the disclosure conversation will take longer and be harder than it needs to be.

What the Notification Must Contain

The notification to the prosecuting office must include: a description of the error and its location in the report, the cases in which the error appears, the agency's assessment of whether the error is material (that is, whether it could affect the outcome of any pending or past proceeding), the scope review findings to date, and the agency's plan for the case-by-case review going forward. It must be in writing, it must be dated and timed, and it must be sent to the specific prosecutor with responsibility for each affected case, not to a general office address.

For cases that have resulted in convictions or pleas where the error may be material, the notification should include the agency's legal counsel's assessment of whether a post-conviction disclosure obligation exists. That assessment should be made jointly with the prosecuting office's own counsel, not unilaterally. The agency's role is to provide the complete factual record; the legal determination of what follows belongs to the joint assessment.

The Defense Disclosure Question

Whether the agency has a direct disclosure obligation to the defense, separate from the prosecutor's Brady obligation, is a question for legal counsel on specific facts. Generally, the Brady obligation runs from the prosecutor to the defense, and the agency's obligation is to provide complete information to the prosecutor. But in cases where the error is discovered post-conviction and the prosecuting office is the same office that tried the case, additional steps may be required to ensure the defendant receives the information. Legal counsel should assess this question for every case in the high-priority tier of the scope review.

The Post-Incident Corrective Record

The incident response is not complete when the immediate disclosure is made and the case-by-case review is underway. The governance board must produce a post-incident corrective record that documents what happened, why it happened, what the scope review found, what was disclosed and to whom, what changes the board made to its policies and procedures as a result, and what monitoring is in place to detect similar errors in the future.

This document serves five audiences. First, the oversight community: civilian review boards, inspector general offices, and city councils will want to know what happened and what the agency is doing about it. A post-incident corrective record that answers those questions completely is better than a fragmented series of responses to individual inquiries. Second, the defense bar: in cases where the error is material, the defense is entitled to a complete account of the error, its scope, and the agency's knowledge of it. Third, future legal proceedings: the corrective record will be discoverable in any civil action or criminal proceeding that arises from the error or the incident response. Fourth, the agency's own operational improvement: the root cause analysis that the corrective record contains should drive changes to the verification standard, the officer training program, and the AI tool's authorized call types. Fifth, the vendor relationship: if the error reflects a model failure, the corrective record documents the basis for a contract remediation conversation.

Root Cause and Systemic Fix

The root cause analysis asks: where did the error originate, and what in the agency's verification, review, and oversight process failed to catch it? The answers are usually not flattering. They tend to fall into one of three categories: the verification standard was not followed (an officer adopted the AI draft without completing the footage-grounded verification pass), the verification standard was followed but was not thorough enough to catch this type of error (the standard needs to be strengthened), or the verification standard was adequate but the error was not recognizable as an error under any reasonable review (the AI tool has a failure mode that requires a technical fix or a call-type exclusion).

Each root cause category requires a different fix. Failure to follow the standard requires a compliance enforcement response and potentially a change to the workflow that makes skipping the standard harder. An inadequate standard requires a policy revision and retraining. A tool failure mode requires a vendor conversation, a call-type exclusion, or a tool replacement, and the governance board's vendor contract should provide the mechanism for that conversation. The corrective record documents which root cause applies and what fix was implemented.

The systemic fix must also address the monitoring gap: how does the agency detect similar errors in the future before they survive fourteen months in a case file? The answer is almost always some combination of a more robust verification standard, a supervisor-level spot review of AI-assisted reports, a quality assurance sampling protocol, and a functioning error-reporting mechanism that reaches the governance board through the standing agenda item described in the previous lesson in this chapter.

When the Error Surfaces Post-Conviction

The hardest version of an AI case problem is the one where the error surfaces after a conviction and the question is whether it affected the outcome. This is the scenario that separates agencies with complete governance records from those without.

An agency with a complete governance record can show: that an AI tool was used, that a specific verification standard was required, that the case record does or does not show that the standard was followed, and that the error in question is of a type that either would or would not have been caught by a properly executed verification pass. That record may not resolve the post-conviction question in the agency's favor. But it is a record the agency can stand behind. It documents that the governance framework existed, that it was designed to prevent exactly this type of error, and that the failure, if there was one, is traceable to a specific step in the process rather than to the absence of any process at all.

An agency without that record faces a much harder question: if there was no verification standard in the charter, no documentation of whether the officer ran a verification pass, and no record of what the AI draft said versus what the officer adopted, then the defense's argument is not that a specific step failed. It is that no step existed. That argument, in a post-conviction context, is the basis for a claim that the conviction rested on an evidentiary foundation that was never validated. It is the argument that the AI did the work and nobody checked.

Key Takeaways

• An AI case problem differs from a routine report error in three ways: the error source may not be traceable, the disclosure obligation triggers immediately at the moment of knowledge under Brady and Giglio, and the scope of the problem is immediately uncertain and may extend across every AI-assisted report the agency has produced.
• The first four hours of incident response require pre-assigned authority and a pre-documented notification chain. Legal counsel assesses the Brady and Giglio disclosure obligation at the moment of first knowledge. Preservation of all original records, including the AI draft before editing, the footage, the CAD entry, and any intermediate versions, happens before any correction is made.
• The scope review identifies the population of AI-assisted reports, conducts an error pattern analysis to determine whether the error reflects a systemic model failure, and produces a risk stratification of the report population that drives the prioritization of case-by-case review.
• Active cases approaching trial are the first priority in the case-by-case review. Cases with recent convictions or pleas where the error could support an appeal are the second priority. The disclosure obligation to the prosecuting office must be fulfilled before any proceeding where the error could be outcome-determinative.
• The notification to the prosecuting office must be in writing, dated and timed, sent to the specific prosecutor with responsibility for each affected case, and must include the error description, affected cases, materiality assessment, scope review findings, and the plan for the ongoing review.
• The post-incident corrective record documents what happened, the scope review findings, what was disclosed and to whom, what policy and procedure changes the board made, and what monitoring is in place going forward. This record serves the oversight community, the defense bar, future legal proceedings, operational improvement, and the vendor relationship.
• The root cause analysis falls into three categories: failure to follow the verification standard, an inadequate verification standard, or a tool failure mode. Each requires a different fix, and the corrective record documents which applies and what was changed.
• A post-conviction AI error is most defensible when the agency has a complete governance record showing the verification standard that existed, whether it was followed in the specific case, and why the error was or was not catchable under a properly executed verification pass. An agency without that record faces the argument that no process existed at all.