The suppression hearing had been scheduled for nine months. The detective had worked the case for two years. Then, three weeks before trial, the defense attorney filed a motion that the prosecution could not ignore: the AI tool used to analyze the interview transcripts had been a version of a commercial large-language model, the agency had not disclosed that fact in discovery, and the defense argued that the outputs of that tool, including a summary the detective had relied on to establish probable cause for the second arrest warrant, constituted material that Brady v. Maryland (1963) required to be produced. The prosecutor had never heard of the tool. The detective had used it routinely for months. Nobody had written a policy about it.

That scenario is not hypothetical speculation about 2035. It is the shape of the legal argument that defense attorneys, public defenders, and civil rights litigators are assembling right now, in 2026, in jurisdictions where agencies have adopted AI tools without building the disclosure architecture to match. The law has not caught up entirely, but the constitutional obligations underneath the law have been settled for sixty years. Brady v. Maryland and Giglio v. United States, the two cases every officer in this lesson must understand before they touch an AI-assisted case file, were decided in 1963 and 1972 respectively. They are not new. The AI is new. The obligation is not.

Brady v. Maryland: The Exculpatory Evidence Floor

In 1963 the Supreme Court decided Brady v. Maryland and established a rule that has governed criminal procedure ever since: the prosecution must disclose to the defense any evidence that is material to guilt or punishment, including evidence that is favorable to the defendant. That category of evidence is called exculpatory evidence: from the Latin "to clear of blame." It is the stuff that makes the defendant look innocent, or less guilty, or that challenges the government's theory of the case.

Brady material does not have to be conclusive. It does not have to guarantee acquittal. It has to be material, meaning there is a reasonable probability that disclosure would have produced a different result at trial. The Supreme Court's test, refined over the decades, asks whether the evidence was suppressed, whether it was favorable to the defense, and whether it was material. If all three conditions are met, the failure to disclose it is a Brady violation, and a Brady violation can vacate a conviction, trigger a new trial, and result in sanctions against the prosecutor who withheld the material. It can also, in its most serious forms, end careers on both sides of the aisle.

Now apply that constitutional rule to AI-assisted investigation. An officer uses an AI tool to summarize twelve hours of interview recordings. The summary is used to draft the affidavit for a search warrant. The tool, by its nature as a generative language model, produced a summary that emphasized the details consistent with guilt. It did not invent facts, but it did omit a portion of the interview in which the suspect described an alibi that the detective ultimately found unpersuasive. The detective did not include the alibi in the affidavit because she had judged it unreliable. The AI tool's summary, which the detective relied on before making that judgment, had also not flagged the alibi. The raw interview transcript was never disclosed.

Is that a Brady problem? The answer depends on whether the alibi information was material, and that is a fact-specific question that will be litigated. But the structural point is clear: when AI assists in the selection, emphasis, and omission of information from a case file, it is participating in a process that has constitutional obligations attached to it. The obligations do not disappear because the selection and emphasis happened inside a model's summarization rather than inside a detective's conscious decision. The disclosure duty runs to the content of the case file, not just to the deliberate choices of the investigating officer.

Brady obligations attach to the evidence, not to the method of analysis. If AI touched the file, the defense may be entitled to know how.

What Counts as a Brady Disclosure in an AI-Assisted Case

The growing practice consensus, though not yet uniform across jurisdictions, is that Brady material in an AI-assisted investigation includes at minimum: the raw source files that the AI analyzed (transcripts, footage, records), the AI-generated outputs that were relied on by investigators (summaries, analyses, flagged patterns), and information about the AI tool itself, including its version, its training data characteristics if known, and any known limitations or failure modes. The last category is the hardest and the most contested, but it is also the most legally interesting as courts begin to develop doctrine around the reliability of AI-assisted evidence.

The practical implication for a patrol officer or detective using AI tools today is this: if you relied on an AI output to make a material decision in the investigation, that output and the underlying source material are candidates for Brady disclosure. You should not be making that determination alone. Your agency needs a written policy, your prosecutor needs to know the tool was used, and the disclosure decision should be made before the defense asks for it, not after.

Prosecutors in a growing number of jurisdictions are requiring exactly this. In 2023, the King County (WA) Prosecuting Attorney's Office took the notable step of barring police reports generated by AI, citing concerns about reliability and the difficulty of disclosure compliance. Other prosecutors have issued guidance requiring that AI-generated content be flagged in discovery packages. The direction of travel is toward more disclosure, not less.

Giglio v. United States: Impeachment Evidence and the Officer as Witness

Nine years after Brady, in 1972, the Supreme Court decided Giglio v. United States and extended the disclosure obligation in a direction that is critically important for officers who testify. The Court held that the prosecution must disclose not just evidence favorable to the defendant on the question of guilt, but also evidence that can be used to impeach (that is, to undermine the credibility of) a government witness. The officer on the stand is a government witness. Evidence that bears on the officer's credibility, including sustained misconduct findings, prior false statements, and material discrepancies between an officer's testimony and documentary evidence, must be disclosed.

In the AI context, Giglio becomes relevant in a specific and practical way. Suppose an officer used an AI tool to draft the narrative section of an incident report. The officer reviewed the draft, made some corrections, and submitted it. The submitted report contains a detail, call it the time a witness statement was taken, that does not appear in the body-worn camera (BWC) footage or the computer-aided dispatch (CAD, the system that logs every dispatch action and timestamp) records. The discrepancy is small, but it is a discrepancy between sworn testimony and contemporaneous records. If the defense discovers it, and if the defense can plausibly argue that the discrepancy originates in an AI hallucination that the officer failed to catch and correct, the Giglio question becomes: does this discrepancy bear on the officer's reliability as a witness?

Courts have not yet resolved how AI-origin discrepancies map onto Giglio doctrine. But the practical risk is clear: an officer who submits an AI-generated report without rigorous verification creates the conditions for a Giglio attack that was preventable. The defense attorney who discovers that the AI tool was used, that the officer acknowledges not remembering what changes were made to the draft, and that the submitted report contains a detail unsupported by the footage has a line of cross-examination that does not require a precedential case to be effective. It only requires the officer to be unable to say, clearly and on the record, "I reviewed every word of this report against the footage and the CAD log, and I verified each factual claim before I signed it."

Giglio applies to officer credibility. An AI-touched report that contains an unverified detail is a Giglio vulnerability the officer could have avoided.

The Brady-Giglio List and AI

Many prosecutors' offices maintain what is informally called a "Brady-Giglio list" (sometimes called a "Giglio list" or "officer credibility file"): a list of officers whose testimony the prosecutor must disclose to the defense because of prior findings that bear on credibility. These lists are built from sustained internal affairs findings, prior false-statement determinations, and in some cases disciplinary records. Being on such a list does not bar an officer from testifying, but it creates an ongoing disclosure obligation that follows every case the officer is involved in.

The concern that emerges from AI-assisted report writing is whether a pattern of reports with unverified AI hallucinations, discovered through defense review, could generate Brady-Giglio material in the form of a systemic credibility question about an officer's review practices. The answer, under current doctrine, is that a single isolated AI discrepancy is unlikely to create Brady-Giglio list exposure. A pattern of discrepancies, especially if the officer cannot describe a consistent verification practice, is a different matter. The discipline of verification, documented and consistent, is the professional protection as much as it is the legal one.

Discovery: The Disclosure Pipeline

Discovery is the legal process by which the prosecution provides the defense with the evidence the government has collected and intends to use, plus the material it is required to disclose under Brady and Giglio and applicable rules. In federal court, Federal Rule of Criminal Procedure 16 governs much of the discovery process. In state courts, the rules vary significantly, but all jurisdictions operate under the Brady-Giglio constitutional floor regardless of their specific procedural rules.

The discovery pipeline in a modern AI-assisted case has several new nodes that did not exist five years ago. First, the AI tool itself: if an officer used a tool like Axon Draft One, which drafts police report narratives from BWC audio and has been reported by testing officers to reduce report-writing time by roughly 82%, that tool's involvement should be documented and disclosed. The tool, the version, the fact of use, and the review process the officer applied to the output are all potentially discoverable. Second, the AI outputs: the draft report before officer edits, the intermediate summaries used during investigation, and the final submitted version with a notation of what was changed are all candidates for the discovery package. Third, the underlying source data: the BWC footage, the CAD log, the interview recordings, and any other digital source material the AI analyzed should be preserved and disclosed as required under standard discovery rules, with no new exceptions for AI-processed versions.

The challenge for agencies right now is that most discovery protocols were written before AI tools entered the evidence-handling pipeline. They cover physical evidence, digital photographs, BWC footage, and records from the records management system (RMS, the agency's central case-file database). They do not always specify what to do with an AI-generated intermediate document, a model output that influenced an investigative decision but was not included in the final report, or a summarization log from a tool embedded in the evidence-management platform. Agencies that have not updated their discovery protocols to address AI-generated material are running a compliance gap that defense attorneys are beginning to exploit.

What the Defense Is Asking For

The defense bar has been paying close attention to AI adoption in law enforcement. Public defenders, civil rights organizations including the Electronic Frontier Foundation (EFF), and defense-side forensic experts have all raised concerns about the transparency and reliability of AI-generated police documents. The EFF has specifically flagged transparency in AI-assisted policing as a civil rights concern, arguing that automated systems that influence charging and prosecution decisions must be subject to independent scrutiny.

In practice, defense attorneys are now routinely including specific discovery requests for AI-related material: (1) any AI tools used in connection with the investigation, the reports, or the case file; (2) any AI-generated drafts, summaries, or analyses; (3) the officer's review log showing what changes were made to AI-generated drafts; (4) the vendor's documentation of the tool's known failure modes and accuracy limitations; and (5) any training the officer received on AI use. These requests are being made in federal court, in state court, and in pre-trial motions that require the prosecution to disclose or object with specificity. Agencies that have no record of the tool used, no log of the review process, and no training documentation are being forced to reconstruct a paper trail under adversarial conditions that is far more difficult and far less favorable than building the documentation from the beginning.

CJIS and Data-Governance Obligations

The CJIS Security Policy (Criminal Justice Information Services, the FBI division that sets security standards for all agencies that access criminal justice information including NCIC, fingerprint databases, and criminal history records) governs how criminal justice data must be handled, stored, and transmitted. CJIS obligations are not optional. Any agency that accesses the national criminal justice information systems must comply with CJIS, and the obligations travel with the data wherever it goes, including into AI tools.

The CJIS concern with AI tools is specific: if an officer feeds case information into a cloud-based AI tool, and that case information includes criminal justice information (CJI, specifically defined by CJIS to include names, dates of birth, arrest records, and criminal history), the cloud tool must be CJIS-compliant. Compliance means the vendor has undergone the required security audit, maintains the required encryption and access controls, and is subject to the required contractual obligations. Not all commercial AI tools are CJIS-compliant. Some tools that are marketed to law enforcement for report writing and summarization are operating in a gray area where the agency has not conducted a formal CJIS compliance determination.

The CJIS obligation stays with the agency, not the vendor. If an officer uses a non-compliant tool to process criminal justice information and there is a data breach or an unauthorized disclosure, the agency bears the compliance failure. The disclosure consequences of a CJIS violation can include losing access to the national criminal justice information systems, which is an operational catastrophe for any law enforcement agency. The CJIS risk is thus not primarily a Brady issue; it is a data-governance and operational issue that must be resolved before an AI tool is deployed, not after it has been in use for a year.

Vendor Contracts and Data Handling

The multi-year, sole-vendor contracts that many agencies are now signing for AI-assisted evidence management and report writing, some on the order of $45 million and up to ten years, typically include data processing addenda that address CJIS compliance. Whether those addenda actually satisfy CJIS is a determination the agency's legal counsel and CJIS systems officer must make, not a matter that can be delegated to the vendor's sales documentation. The lock-in structure of these contracts, where the camera, drone, cloud storage, and AI report-drafting tool are all bundled from a single vendor, means that a CJIS compliance gap discovered two years into a ten-year contract is expensive and operationally disruptive to remediate.

Officers do not need to become CJIS experts. They do need to know that the tool they are using has been vetted by their agency for CJIS compliance, and they need to ask the question if no one has told them. The question "has this tool been cleared for CJIS-compliant use?" is a reasonable professional question before feeding a case file into any AI platform, and an agency that has no answer to that question has a governance gap that needs to close.

The Suppression Motion and What It Costs

A suppression motion (formally, a motion to suppress) is a defense motion asking the court to exclude evidence because it was obtained in violation of the defendant's constitutional rights, most commonly the Fourth Amendment (unreasonable search and seizure), the Fifth Amendment (compelled self-incrimination), or the Sixth Amendment (right to counsel). If the court grants the suppression motion, the evidence cannot be used at trial, and in many cases the loss of that evidence effectively ends the prosecution.

Brady-based suppression motions are a distinct and serious category. The defense argues that the prosecution withheld material exculpatory or impeachment evidence, and asks the court to suppress the evidence obtained from the tainted investigation, dismiss the charges, or declare a mistrial. In a case where the prosecution's failure to disclose AI involvement is the basis for the motion, the remedy the court imposes depends on whether the failure was willful, reckless, or inadvertent, and on what prejudice the defendant suffered.

Consider the practical cost of a successful suppression motion in an AI disclosure case. The officer spent two years building the case. The agency spent, conservatively, a hundred thousand dollars in labor costs on the investigation. The AI tool was used legitimately to manage the volume of interview recordings. Nobody documented the tool's use because no policy required it. The defense discovers the tool was used through a FOIA (Freedom of Information Act) request for the officer's work records and email. The suppression motion is filed at the six-month mark before trial. The court holds an evidentiary hearing on whether the AI summary influenced the probable cause determination for the second warrant. The motion is not granted, but the hearing costs three weeks of prosecutor time, requires the detective to testify about the AI tool's use, and reveals to the defense exactly which portions of the AI output the detective relied on. The case goes to trial with a weakened prosecution narrative and a defense theory built around the AI tool's reliability. A case that was strong nine months ago is now a risk.

The point is not that AI use automatically generates suppression problems. The point is that undisclosed AI use, in an environment where defense attorneys know to ask about it, creates litigation risk that proper disclosure policies would largely eliminate. Transparency about AI tool use, properly documented, is not a liability in the courtroom. It is a credibility asset. The officer who can say, "Yes, I used this tool. Here is what it produced. Here is the review I conducted. Here is what I changed. Here is what I verified against the footage. And here is the final report I adopted as my sworn account," is a far stronger witness than an officer who is reconstructing that story under cross-examination after the defense has already found the tool in the email records.

Building the Disclosure Record Before the Defense Asks

The practical question for every officer and every agency is: what does good AI disclosure practice look like right now, in 2026, before the courts have fully resolved the doctrine?

The emerging consensus from prosecutors who have addressed the issue, from the King County ban, from EFF's published concerns, and from the first wave of discovery disputes, points toward six elements of a disclosure-ready AI practice.

Document the tool. When an officer uses an AI tool in connection with a case, the tool, the version, and the date of use should be recorded in the case file. This is a two-sentence notation in the RMS, not a dissertation. "Axon Draft One v2.3 used to draft initial narrative on [date]. Officer reviewed draft against BWC footage and CAD log before submission."

Preserve the AI output. The initial AI-generated draft, before officer edits, should be retained in the case file or the evidence management system. If the prosecution later needs to demonstrate what the AI produced and what the officer changed, the original draft is the evidence. Deleting AI drafts after submitting a modified version creates exactly the discovery gap that suppression motions exploit.

Document the review. The officer's verification process, what was checked and against what source, should be logged. A brief verification record: "Checked time of initial contact against BWC timestamp: matches. Checked description of subject clothing against BWC still capture: matches. Checked witness name spelling against CAD log: corrected in report," creates a contemporaneous paper trail that supports the officer's testimony and satisfies the prosecution's disclosure obligation.

Disclose to the prosecutor early. Prosecutors need to know about AI tool use before they prepare their discovery packages, not when the defense motion lands. The agency's disclosure policy should require officers to flag AI-assisted case files to the assigned prosecutor, and the prosecutor's office should have a protocol for deciding what to include in the Brady and Giglio disclosure package in AI-touched cases.

Follow agency and CJIS policy. The use of AI tools in case investigation should be governed by an agency-level use policy that has been reviewed by legal counsel and approved by command staff. Officers using tools outside an approved policy are creating personal exposure, not just agency exposure. If the agency has no policy, the officer should ask. If the answer is "we have not thought about that," the officer should escalate the question. The absence of a policy is not permission.

Be ready to testify about it. Every officer who uses an AI tool in connection with a case should be able to explain, clearly and on the stand, what the tool did, what they reviewed, and what they verified. If an officer cannot articulate that explanation, they are not yet ready to use the tool in a way that will hold up in court. The preparation for the deposition question, "Officer, did you write this report, or did a computer?", begins not in the witness box but in the verification step after the AI draft is generated.

Key Takeaways

• Brady v. Maryland (1963) requires the prosecution to disclose exculpatory evidence: material favorable to the defense that, if disclosed, creates a reasonable probability of a different result. When AI tools assist in case investigation, the outputs of those tools and the source material they analyzed are candidates for Brady disclosure.
• Giglio v. United States (1972) extends the disclosure obligation to impeachment evidence about government witnesses, including officers. An AI-generated report with unverified details that contradict the footage or the CAD log creates Giglio vulnerability that rigorous officer verification would have prevented.
• Discovery, the legal process for disclosure of government evidence, must now address AI-generated intermediate documents, summaries, and analysis logs that influenced investigative decisions. Most agency discovery protocols have not yet been updated to cover this material, and the gap is being exploited in suppression motions.
• The CJIS (Criminal Justice Information Services) Security Policy requires that any AI tool processing criminal justice information be compliant with CJIS security standards. The compliance obligation stays with the agency. Officers should confirm CJIS clearance before feeding case files into any AI platform.
• A suppression motion based on undisclosed AI involvement costs far more in litigation time, prosecutor resources, and case credibility than the disclosure would have cost at the outset. Transparency about AI use, properly documented, is a credibility asset in court, not a liability.
• The King County (WA) Prosecuting Attorney's Office barred AI-written police reports; the EFF has raised transparency concerns about AI-assisted policing. These positions represent real governance pressures that officers and agencies need to know and take seriously.
• Six elements of a disclosure-ready AI practice: document the tool and version in the case file, preserve the original AI draft, log the verification steps taken, disclose to the prosecutor early, follow agency and CJIS policy, and be prepared to testify about the process clearly and on the record.
• Brady and Giglio obligations are constitutional, settled law. They predate AI by decades. The officer's job is not to wait for the case law to catch up but to apply those existing obligations to the new tool with the same professional discipline they apply to every other element of the case file.