The defense attorney's letter arrived on a Tuesday. She represented a defendant whose case had been built substantially on an AI-drafted incident report, a body-worn camera (BWC) summary, and a case chronology assembled with AI assistance at every stage. Her letter demanded, under Brady v. Maryland (the 1963 Supreme Court decision requiring prosecutors to disclose material exculpatory evidence to the defense), the original AI-generated drafts, any prompts used to generate them, any corrections the officer made to those drafts, any outputs the AI produced that were rejected and not included in the final report, and the full disclosure of what the AI tool was, how it worked, and what data it had access to. The prosecutor called the detective on the case, who called the records unit supervisor, who called the department's legal counsel. The question was whether the agency could assemble what the defense had requested, document what the AI had done and how the officer had reviewed it, and produce a package that satisfied the Brady obligation without opening a suppression motion. The answer was yes, but only because the officer had followed the disclosure-by-design workflow from the day the report was filed.

Brady, Giglio, and What They Mean for AI-Touched Files

Brady v. Maryland establishes that the prosecution must disclose material exculpatory evidence to the defense. Giglio v. United States (1972) extends this to impeachment evidence, meaning evidence that could be used to challenge the credibility of a prosecution witness, including a police officer. Together, Brady and Giglio create what practitioners call the Brady/Giglio disclosure obligation, and together they frame the entire question of AI use in public safety as a constitutional matter, not just an administrative one.

When AI touched the file at any stage, the Brady/Giglio analysis must ask: could any AI-generated content, any AI-produced gap-fill, any content rejected by the reviewer but retained in the platform's version history, or any fact about how the AI tool works be "material" to the defense? Under the Brady standard, evidence is material if there is a reasonable probability that disclosure would produce a different outcome. That is a broad standard. A gap-fill detail that the officer corrected before the report was final might be material: it shows that the AI initially described the incident differently than the officer ultimately reported, and a defense attorney can argue that the AI's initial description, based on the same audio, is itself evidence about what the recording actually contained.

Giglio is equally important. If an officer's credibility is at issue, including an officer's credibility about how they drafted a report, evidence that the report was AI-assisted and that the officer did not follow the required verification protocol could be impeachment evidence. This is one reason the King County (Washington) prosecutor's office barred AI-written police reports: prosecutors who cannot fully account for the genesis of a report they are relying on cannot confidently assert the report's integrity under Brady/Giglio scrutiny. The ban was not an indictment of the technology. It was a recognition that without a documented, auditable disclosure standard, the Brady obligation became unmeetable.

When AI touched the file, Brady/Giglio does not go away. The disclosure obligation expands to include the AI process: what the tool produced, what the officer accepted or rejected, how the review was conducted, and what the tool had access to. The disclosure package must answer all of it.

The EFF Transparency Concern and Why It Matters Operationally

The Electronic Frontier Foundation (EFF), a civil liberties organization that monitors law enforcement technology, has raised specific transparency concerns about AI police report writing. The EFF's core concern is that AI-generated police reports may systematically incorporate language from historical reports, including reports that may have encoded biased or legally problematic framings. If an AI model is trained on a corpus of police reports from an agency with documented use-of-force patterns, it will reproduce those patterns in its outputs. A defense attorney who knows this can argue that every AI-assisted report in the agency is tainted by the statistical regularities of the training corpus, and that the report in their case may reflect what "usually happens" in similar incidents at this agency rather than what actually happened in this incident.

The EFF's concern is not merely theoretical. It creates a concrete operational obligation: agencies that use AI-assisted report writing must be able to explain, specifically and completely, what the AI tool was trained on, what guardrails it operates under, what its known failure modes are, and what verification protocol was applied in this specific case. If the agency cannot answer those questions, it cannot meet the Brady/Giglio transparency demand, and it hands the defense an argument for suppression.

The practical implication is that the disclosure package for any AI-touched file must include not just the outputs of the AI tool but adequate documentation of the tool itself. This is new ground for most records units. The standard disclosure package for a pre-AI case included the incident report, supplemental reports, CAD entry, RMS entries, body camera footage, and related evidence. The AI-era disclosure package includes all of that plus the AI disclosure documentation. What that documentation contains is what the rest of this lesson explains.

What a Complete AI Disclosure Package Contains

A disclosure package for a case involving AI-assisted documentation has six components. Understanding each component, what it contains and why it is required, is the foundation for building a disclosure-by-design workflow.

Component One: The Adoption Record

The adoption record documents that a human officer reviewed, corrected, and adopted the AI draft as their own sworn account. It should include the officer's name and badge number, the date and time of review, a statement that the officer reviewed the AI-generated draft against the BWC recording and other available records, a statement of any corrections made, and the officer's electronic signature or other form of attestation.

The adoption record is the foundational document in the disclosure package. Its purpose is to demonstrate, specifically and traceably, that the report is not an unreviewed AI output but a sworn account adopted by a named officer who is legally responsible for its accuracy. Without the adoption record, the defense can argue that the report was generated by a machine, reviewed perfunctorily or not at all, and submitted as a sworn statement in violation of the officer's duty to personally attest to the accuracy of their reports. With the adoption record, the defense knows exactly who reviewed what, when, and what they changed.

Component Two: The Version History

The version history is the record of what the AI generated versus what the officer submitted as the final report. It includes the AI-generated initial draft (or a summary of the AI-generated content where the full text is too voluminous), the specific changes the officer made during the review, and the final adopted report. Many AI-assisted report platforms retain this version history automatically in the platform's audit log. The disclosure package must include it, in a format that allows the defense to compare what the AI produced to what the officer submitted.

The version history is a Brady document. If the AI's initial draft described the incident differently than the officer's final report, that difference is potentially material. Consider the detective whose AI draft said "both hands raised" when the footage showed one hand raised: the AI's initial description is a record of what the model inferred from the audio, and it is different from the officer's corrected version. The defense can argue that the AI's inference and the officer's observation are both relevant to what the footage actually showed. That argument may not succeed, but the defense is entitled to make it, which means they are entitled to the version history that makes it possible.

Component Three: The Verification Log

The verification log documents the footage-grounded verification pass: what the officer checked, against which source (BWC footage, CAD entry, field notes), and with what result. A complete verification log should include at minimum the timestamp ranges reviewed, whether any discrepancies were found and what they were, how discrepancies were resolved, and a statement that the officer completed the verification pass before adopting the report.

The verification log is the Brady/Giglio answer to the question: how do we know the report is accurate? The officer's answer in a deposition should be: "I ran the footage-grounded verification pass, documented in the verification log, and I can show you exactly what I checked, what I found, and what I changed." That answer, backed by the log, is a complete and auditable account of the officer's due diligence. Without the log, the officer is left with the unsupported assertion that they "reviewed" the report, which a skilled attorney can challenge effectively.

Component Four: The Tool Disclosure

The tool disclosure is a standardized document, prepared at the agency level not the individual case level, that describes the AI tool used in report writing: its name, version, the vendor, a summary of how it works (draft from BWC audio, draft from interview transcript, case summarization, or other function), the training data policy as disclosed by the vendor, the known limitations of the tool as disclosed by the vendor, and the agency's verification protocol that applies to all AI-drafted reports.

The tool disclosure is prepared once per tool deployment and updated when the tool changes materially. It is incorporated by reference into every disclosure package for cases involving that tool. This means the detective does not write a custom description of the AI tool for every case. The agency maintains a current, accurate tool disclosure document, and every disclosure package points to it.

The tool disclosure must cover the CJIS (Criminal Justice Information Services) compliance status of the tool: whether the vendor has a current data-use agreement with the agency, whether the processing infrastructure meets CJIS Security Policy requirements, and how long the vendor retains data processed through the platform. These are not details a defense attorney needs to see in every disclosure package, but they must be part of the agency's documentation and available on request. If a defense attorney argues that data from the case was processed by a vendor that retains it and that the vendor's data set may have been used to train the model, the agency needs a documented answer.

Component Five: The Redaction Log for Released Material

If the disclosure package includes footage or written records that have been redacted, the redaction log is a mandatory component. For criminal case disclosure under Brady/Giglio, the redaction standards are different from public-records release: the defense generally has a right to material relevant to the case even if that material would be withheld from a public-records requester. Third-party faces and license plates in footage are still typically redacted in criminal case disclosure, but the officer's face is not, and the basis for any redaction must be specifically documented because a defense attorney can challenge each one.

The redaction log in the criminal case disclosure context should state, for each redaction: what category of information was redacted, the specific legal or policy basis for the redaction, and whether the redacted information is relevant to the defendant's case and, if so, why it was nonetheless redacted. The last element is the most important: a redaction that removes potentially relevant material from a criminal case disclosure is a Brady problem unless the basis for the redaction is legally sound and thoroughly documented.

Component Six: The Disclosure Cover Letter

The disclosure cover letter is the entry point into the package. It lists every document being disclosed, identifies any documents being withheld and the specific legal basis for each withholding, states that the case involved AI-assisted documentation, specifies which tool and describes the verification protocol, and provides the name and contact information of the responsible records officer who can answer questions about the package.

The cover letter is the first document the defense attorney reads. If it is incomplete, unclear, or fails to identify AI use, the defense attorney has grounds to argue that the agency's disclosure was deficient and to seek additional disclosure, including potentially an in camera (private, judge-only) review of materials the agency claims are withheld for valid reasons. A complete, accurate cover letter is the agency's first and best defense against a successful motion for additional disclosure.

Building Disclosure by Design into the Workflow

The six components of the disclosure package do not materialize at the time the defense makes a Brady demand. They must be built into the workflow from the moment the AI tool is used. This is what "disclosure by design" means: the documentation is created contemporaneously with the work, not reconstructed after a demand arrives.

Consider what happens when disclosure documentation is not built in. The detective in the opening scenario, if they had not followed the disclosure-by-design workflow, would have faced the defense attorney's Brady demand with no adoption record, no version history, no verification log, and no tool disclosure document. The detective would have been required to reconstruct these documents from memory, from whatever the AI platform retained in its audit log, and from the officer's personal notes, which may not exist. The reconstruction would have been incomplete, the incompleteness would have generated a motion, the motion would have consumed prosecution resources, and the outcome would have been uncertain. The correct workflow, followed from day one, makes the demand answerable in hours rather than weeks.

The adoption record is created when the officer adopts the report: it is the final step in the report-submission workflow, and the platform should require it before submission is possible. The version history is created automatically by the platform if it retains version history, and the agency must ensure this retention is configured. The verification log is created during the verification pass: the officer logs it as they work, and the platform should provide a structured verification log field rather than relying on freeform notes. The tool disclosure is a standing document maintained by the agency and referenced in every disclosure package. The redaction log is created during the redaction review. The cover letter is prepared at the time of disclosure.

None of these steps require significant additional time if they are built into the workflow. The adoption record takes two minutes if there is a standard form. The verification log takes the time of the verification pass, which is time the officer is spending anyway. The tool disclosure is done once. The redaction log is generated by the platform. The cover letter is templated. The total added documentation burden, over a well-designed workflow, is measured in minutes per case, not hours. The benefit is measured in months of avoided litigation.

The Bundled Contract Problem and Vendor Accountability

An important and frequently underappreciated dimension of the disclosure package involves the vendor's role in creating it. Many law enforcement agencies have entered into bundled, multi-year, sole-vendor contracts covering cameras, drones, cloud storage, and AI tools. Contracts on the order of $45 million over ten years are not unusual, and the lock-in they create is substantial. When a single vendor provides the BWC platform, the AI drafting tool, the cloud storage, the evidence management system, and the CAD integration, the disclosure package depends on what that vendor's systems capture and retain.

If the vendor's platform retains version history for ninety days and the case goes to trial in fourteen months, the version history is gone. If the vendor's audit log does not capture the specific timestamp of each verification pass, the verification log is incomplete. If the vendor does not provide exportable, human-readable records in a format courts and defense attorneys can review, the disclosure package cannot be assembled from the platform's data alone.

This means the agency's procurement and contract requirements must include specific documentation about retention periods, export formats, audit log content, and data availability for litigation purposes. The agency that signs a ten-year, bundled contract without these requirements has given a vendor control over its disclosure obligations. When the defense attorney asks for the version history and the vendor's platform does not retain it, the problem is not the vendor's problem. It is the agency's Brady problem.

The EFF has identified vendor lock-in and opacity as specific concerns: if the agency cannot describe how the AI tool works, what it was trained on, and what it retains, it cannot meet the transparency obligations that Brady/Giglio impose. Agencies negotiating or renegotiating contracts with bundled AI vendors should treat documentation retention and audit log completeness as non-negotiable contract requirements, not optional features.

When the Package Is Challenged: The Suppression Motion Scenario

The most serious challenge to an AI-touched disclosure package is a suppression motion: a defense motion arguing that evidence should be excluded from trial because it was obtained in violation of the defendant's constitutional rights. A suppression motion based on Brady grounds argues that the prosecution failed to disclose material evidence and that the failure was prejudicial.

The scenario that testing officers and prosecutors most fear is the following. A defendant is convicted. On appeal, the defense discovers that the AI-assisted report contained a gap-fill detail that was not caught by the verification pass, that the gap-fill mischaracterized a critical element of the defendant's conduct, and that the disclosure package did not include the AI's initial draft. The appellate court, finding a Brady violation, vacates the conviction. The defendant is retried or released. The officer's report is now permanently associated with a documented AI failure that was not disclosed.

This scenario is not hypothetical in structure. It is the logical downstream consequence of an under-documented AI-assisted report workflow. The protection against it is complete disclosure by design: the verification log that shows the officer checked the gap-fill region, the version history that shows the officer corrected the draft to reflect the footage, and the adoption record that shows the officer attested to the accuracy of the corrected version. If all three exist, the suppression motion fails because the Brady obligation was met. If any one is missing, the motion has traction.

Prosecutors in jurisdictions where AI-assisted reporting is deployed should establish a pre-trial disclosure checklist specifically for AI-touched cases. The checklist should include: adoption record present, version history present, verification log present, tool disclosure present, redaction log present (where applicable), and cover letter complete. A case that reaches trial without all five is a case with an open Brady flank that the defense will eventually find.

Key Takeaways

• Brady v. Maryland and Giglio v. United States together make AI use in law enforcement a constitutional disclosure matter. When AI touched the file, the defense is entitled to the AI's outputs, the officer's corrections, the verification record, and adequate documentation of what the tool was and how it worked.
• The King County (WA) prosecutor's bar on AI-written police reports reflects the difficulty of meeting Brady/Giglio obligations without a documented, auditable disclosure standard. The solution is not to ban the tool but to build the disclosure standard into the workflow.
• The EFF's transparency concern is operationally real: an agency that cannot explain what the AI was trained on, what its known failure modes are, and what verification protocol was applied cannot meet the transparency demand of the Brady/Giglio inquiry.
• A complete AI-era disclosure package has six components: the adoption record, the version history, the verification log, the tool disclosure, the redaction log (where applicable), and the disclosure cover letter. Each must be created contemporaneously, not reconstructed after a Brady demand.
• Bundled, multi-year, sole-vendor contracts (on the order of $45M and up to 10 years) can give vendors control over the retention and format of the documentation needed for Brady disclosure. Retention periods, export formats, and audit log content must be specified as contract requirements before signing.
• Suppression motions based on Brady grounds are the downstream consequence of under-documented AI workflows. The version history showing the AI's initial draft and the officer's corrections is a Brady document: if the AI described the incident differently than the officer's final report, the defense is entitled to know.
• Disclosure by design means building the documentation into the workflow as the work happens. The total added burden is minutes per case in a well-designed system. The alternative is weeks of litigation and an uncertain outcome.
• Prosecutors handling AI-touched cases should maintain a pre-trial Brady checklist: adoption record, version history, verification log, tool disclosure, redaction log, and cover letter. A case missing any one of these has an open constitutional disclosure flank.