The union rep got a call from Officer Chen at 9 p.m. on a Tuesday. Chen had just come out of a meeting with his lieutenant about a report that was under review. A defense attorney had filed a motion arguing that a quote in Chen's report, attributed to the defendant, did not match the body-worn camera footage. The lieutenant wanted to know how the report had been produced. Chen had used the department's new AI drafting tool. He had reviewed the draft. He was confident the report was accurate overall. But the lieutenant was asking questions Chen could not fully answer: "What is a hallucination?" "What does Brady require?" "Is this a disclosure issue?" "Does the agency need to log this?" Chen had heard these words in training. He had not learned them in a way that connected to his job. He was not sure what any of them meant for the situation he was actually in. This lesson is the one that would have helped Officer Chen. Every term here is defined in field language, grounded in the specific scenario where it matters, and tied to a consequence: a legal consequence, a career consequence, or a case-outcome consequence. These are not vocabulary words. They are the tools you need to operate AI responsibly in public safety work, and to explain that operation to anyone who asks.

AI Tool Terms: The Ones in the Field

Before getting to legal and evidentiary terms, it helps to have a shared vocabulary for the AI tools themselves, because the same word is used in very different ways in different conversations. Here are the terms you will encounter in the field and what they actually mean.

Large Language Model (LLM)

A large language model is an AI system trained on enormous amounts of text to predict and generate language. GPT, the technology behind many AI writing assistants, is an LLM. The AI drafting tools appearing in body-camera platforms and report-writing systems are typically built on or closely related to LLMs. The key characteristic for public safety work: LLMs generate text based on statistical patterns. They do not retrieve and verify facts. They produce what is probable and coherent, which is usually close to accurate but not guaranteed to be.

Generative AI

Generative AI refers to AI systems that produce new content, text, images, audio, or other media, rather than simply analyzing or classifying existing content. Report-drafting tools, interview summarizers, and translation assists in dispatch are all generative AI. The term matters because it distinguishes these tools from older AI or algorithmic tools that only sort, rank, or flag. Generative AI creates, and what it creates can contain hallucinations.

Hallucination

Hallucination is the term used when an AI model produces output that is internally coherent and confident in tone but factually wrong, unsupported by the source material, or fabricated. The word can be misleading because it suggests the AI is experiencing something like a sensory error. A more precise description is: the model generated text that is statistically probable and contextually appropriate but does not accurately reflect the source material it was given.

In a public safety context, a hallucination in a police report is a statement, a detail, a quote, or a sequence that the AI generated but that is not in the body-camera footage, the CAD entry, or any other verified source. The danger is that hallucinations are tonally indistinguishable from accurate extractions. The AI does not flag them. They look exactly like everything else in the draft. Officer Chen's situation, a quote that did not appear in the footage, is a hallucination of the invented-quote type.

Grounding and Context Window

Grounding refers to the practice of giving an AI model specific source material, the footage transcript, the CAD entry, the incident notes, and instructing it to generate output based only on that material. A grounded model is less likely to hallucinate than an ungrounded one because it has specific content to work from rather than generating from general patterns alone.

A context window is the block of text the model can "see" at one time when generating output. If the report-drafting tool gives the model a transcript of the BWC (body-worn camera) audio, that transcript is in the context window. If the transcript is longer than the context window, the model may not have access to the full recording when generating the early or late sections of the report. Long incidents with a lot of footage can push important content outside the context window, which is one reason long incidents are higher-risk for hallucination.

Prompt and System Prompt

A prompt is the instruction given to an AI model. "Draft a police report narrative from this transcript" is a prompt. The quality and specificity of the prompt affects the quality of the output. A prompt that says "draft only what is in the transcript and flag anything that is unclear" will produce different output than a prompt that says "draft a complete report."

A system prompt is a hidden instruction that sets the model's behavior for a specific deployment. In a report-drafting tool, the system prompt might instruct the model to draft only from verified source material, to flag gaps, and to use specific report format conventions. Individual users typically cannot see the system prompt. Understanding that system prompts exist explains why the same model can behave very differently in different tools: the tool's system prompt shapes the output in ways that the user's own prompt does not.

Legal and Evidentiary Terms: Field Definitions

These are the terms that live at the intersection of AI tools and the legal consequences of using them. Every officer, dispatcher, and records staff member who works with AI tools needs to understand these terms in the specific context of their work.

Brady Material, Brady Obligation

Brady refers to Brady v. Maryland, the 1963 Supreme Court decision requiring the prosecution to disclose any evidence that is material to guilt or punishment and favorable to the defendant. This type of evidence is called Brady material or exculpatory evidence. The disclosure requirement is constitutional, grounded in due process, and is not optional even if the prosecution judges the evidence to be unhelpful to its case.

In the AI context, Brady is relevant in two directions. First: if an AI-drafted report contains a hallucinated incriminating detail that is not in the footage, and the defense obtains the footage and finds the discrepancy, the footage is Brady material. The prosecution must disclose the footage, which contradicts its own report. Second: if the AI draft softened a fact that was favorable to the defense, and the softened version appears in the report while the unambiguous version is in the footage, the footage again is Brady material. The ground truth is the footage. Anything the AI did to change that ground truth in the report creates a disclosure obligation.

Practical field meaning: before you adopt an AI-drafted report as your sworn account, ask yourself whether everything in the report is consistent with the footage. If it is not, the footage is Brady material and the prosecution must know about the discrepancy before the case proceeds.

Giglio Material, Giglio Obligation

Giglio refers to Giglio v. United States, the 1972 Supreme Court decision extending Brady's disclosure requirements to impeachment evidence, specifically evidence that bears on the credibility of a witness. For police officers who testify, Giglio material includes prior findings of untruthfulness, prior disciplinary actions for report inaccuracies, and any documented pattern of conduct that would help the defense challenge the officer's credibility.

In the AI context, Giglio has a specific and important implication: if an officer has a documented history of adopting AI-generated drafts that were found to contain inaccuracies, that history is Giglio material. The defense is entitled to know about it before the officer testifies. And the prosecutor's office has an obligation to disclose it. This is why agencies need accurate documentation of AI use and verification passes: not just to prove that verification happened, but to create an accurate record of what the officer's verification practice actually was.

Practical field meaning: the way you handle AI drafts today creates a record that can be part of your credibility as a witness in future cases. A documented, consistent verification practice is a credibility asset. A documented pattern of cursory adoption is a credibility liability.

Disclosure and Discovery

Disclosure and discovery are related but distinct concepts that come up constantly in discussions of AI in public safety. Disclosure, in the Brady-Giglio sense, is the prosecution's affirmative obligation to provide favorable evidence to the defense. It is proactive: the prosecution must identify and disclose Brady and Giglio material without waiting to be asked.

Discovery is the pre-trial process by which both sides exchange evidence and information. Defense counsel can request, through discovery, any material that is relevant to the case, including documents that show how a police report was produced. If the defense knows the report was AI-assisted, it can request the AI draft, the tool's logs, the agency's AI policy, and documentation of the verification process. If the agency has those things, it produces them. If it does not, the absence of documentation becomes a finding that can affect the case.

Practical field meaning: AI use in report production is discoverable. Agencies and officers should operate on the assumption that their AI workflow, their verification practices, and their documentation will be examined in any case that goes to litigation. The way to make that examination comfortable rather than damaging is to have a policy, follow it, and document that you followed it.

Chain of Custody

Chain of custody is the documented, unbroken record of who possessed an item of evidence, when, and what was done with it. It is most commonly associated with physical evidence: a gun collected at a crime scene has a chain-of-custody log that tracks every transfer from scene to evidence vault to laboratory to courtroom. Any break in the chain is a potential challenge to the evidence's integrity.

In the AI context, chain of custody is expanding to cover documentary evidence. A police report produced with AI assistance has a chain of production: the footage was captured, the audio was processed by the AI tool, a draft was generated, the officer reviewed it, the officer made corrections, and the officer submitted the final report. If any of these steps is undocumented, the defense can argue that the chain of custody for the report's content is broken, that content was introduced into the document at an unverified step. An agency that documents each step, including what the AI generated and what the officer changed, has an intact chain of custody. An agency that produces only the final report, with no documentation of the AI step or the verification process, has a gap in the chain.

Admissibility

Admissibility refers to whether a piece of evidence is legally acceptable for consideration by a court. Evidence that is inadmissible cannot be presented to the jury. Admissibility is governed by rules of evidence, which vary by jurisdiction and case type, and it can be challenged by either party through a motion to suppress or an objection at trial.

AI-assisted reports are not inherently inadmissible. But the path to admissibility runs through authenticity: the report must be what it claims to be, an accurate sworn account of the officer's observations. A report that contains AI-generated content that was not verified against the footage and that contains inaccuracies faces admissibility challenges. Not because it was AI-assisted, but because a document containing inaccurate sworn statements is a document whose accuracy can be challenged. The tool that produced the inaccuracy does not change the standard; it just adds a layer to the story of how the inaccuracy got there.

Records and Operations Terms

The following terms appear in the records, dispatch, and operations contexts where AI is also active, and where the legal stakes, while sometimes lower than in a criminal trial, are still significant.

Redaction

Redaction is the process of removing or obscuring sensitive information from a document or recording before it is released or disclosed. In public safety, redaction is required when footage or documents contain personal identifying information about third parties, information that would compromise an ongoing investigation, or information exempt from disclosure under open-records statutes. AI-assisted redaction automates the initial identification of content that should be redacted: faces in footage, license plates, medical information, names of minors.

AI redaction tools are useful for reducing the hours a records unit spends on release review. But they are not infallible. They miss things: a face partially obscured by a shadow, a name visible on a piece of paper in the background, a license plate seen at an angle. The human reviewer who follows the AI pass is the quality-control step. Over-redaction (blurring too much, withholding information that should be released) and under-redaction (missing something that should have been redacted) are both failures. Over-redaction can produce open-records complaints and legal challenges from requesters. Under-redaction is a privacy violation and can produce civil liability for the agency.

CJIS (Criminal Justice Information Services)

CJIS, the Criminal Justice Information Services, is the FBI division that maintains the national criminal justice databases, including the National Crime Information Center (NCIC), and publishes the CJIS Security Policy that governs how criminal justice information (CJI) is handled, stored, and transmitted by agencies that access these systems. The CJIS Security Policy sets minimum standards for access control, audit logging, encryption, personnel security, and physical security for any system that handles CJI.

When AI tools process body-camera footage, interview transcripts, RMS records, or other material that is or contains CJI, the agency's CJIS obligations apply to that processing. The vendor who processes the data on the agency's behalf must meet CJIS requirements. The agency must verify that compliance before deploying the tool. Using a consumer AI tool, a personal account, or an unofficial cloud service to process CJI is a CJIS violation. The obligation stays with the agency, not the vendor, even when the vendor is processing the data.

Practical field meaning: if you have case material, footage, interview notes, RMS records, or any other criminal justice information, you may only process it through agency-approved, CJIS-compliant tools. Any other path is a policy violation with potential consequences for the case and for the officer.

CAD and RMS

Computer-Aided Dispatch (CAD) is the system dispatchers use to receive, log, and manage 911 calls and dispatch resources. Every call generates a CAD entry, a record of the incident type, address, time, units dispatched, and notes. The CAD entry is part of the evidentiary record for any call that results in an incident report.

Records Management System (RMS) is the system agencies use to store and manage case reports, incident records, arrest records, evidence logs, and related documentation. The RMS is where police reports live after they are submitted. It is also the source from which reports are pulled for prosecution, discovery, and public records requests.

AI touches both systems in current deployments: AI can assist CAD call classification and entry, and AI report-drafting tools typically push completed drafts into the RMS workflow. Understanding these systems as the repositories of AI-assisted output helps frame the scope of the AI governance challenge: errors that enter the CAD entry or the RMS report do not stay in one place, they propagate through every system and person who accesses those records.

BWC (Body-Worn Camera)

A Body-Worn Camera (BWC) is the recording device an officer wears during patrol and calls. The BWC footage is the primary source material for AI report-drafting tools: the draft is generated from the footage's audio and, in some systems, its visual content. The BWC footage is also the verification standard: the footage is the ground truth against which every statement in the AI-drafted report must be checked.

The BWC's evidentiary role means that the relationship between the footage and the report is legally significant. A report that accurately reflects the footage is a consistent, strong evidentiary record. A report that diverges from the footage, whether because of AI hallucination, officer error, or any other reason, creates an inconsistency that the defense will surface. The footage travels with the case through discovery, and its consistency with the report is examined by prosecutors, defense attorneys, and courts.

Putting the Vocabulary to Work: Chen Revisited

Let's return to Officer Chen and apply this vocabulary to his situation. Chen used a generative AI tool, built on an LLM, to draft a report from his BWC footage. The tool produced a draft that included an invented quote, a hallucination of the invented-quote type. Chen reviewed the draft and adopted it as his sworn report without verifying every statement against the footage, so the hallucinated quote survived into the final report.

The defense attorney, who obtained the BWC footage through discovery, found the discrepancy: the quote in the report did not appear in the footage. The footage is now Brady material: it is exculpatory (or at least inconsistent with the prosecution's document), and the prosecution must disclose and address the inconsistency. Chen's verification practice in this case, and in prior cases, is now a Giglio concern for the prosecution: his pattern of verification is relevant to his credibility as a witness.

If Chen's agency had a documented AI use policy, a verification checklist, and a log of what Chen verified before adoption, the agency could demonstrate that the error was a single instance, that the verification process was followed in general, and that the specific hallucination survived despite a reasonable review. That is a defensible position. If the agency has no policy, no checklist, and no log, then the question of whether Chen ever verifies anything in his AI drafts is wide open, and the defense will use that opening.

The terms in this lesson are the tools Chen needed. He needed to know what a hallucination was and why it would not announce itself. He needed to know what Brady required and that the footage would be disclosed. He needed to know what Giglio meant for his credibility record. He needed to know what chain of custody meant for the report's provenance. He needed to know what CJIS required for the tool he was using. And he needed to know that as the adopting officer, he was the author, accountable for every word, regardless of how the draft was produced.

Key Takeaways

• Hallucination is AI output that is coherent and confident but factually wrong, unsupported, or fabricated. It does not announce itself. Hallucinations look identical to accurate AI output. The only detection mechanism is human verification against the source material.
• Brady v. Maryland (1963) requires the prosecution to disclose exculpatory evidence. An AI hallucination that creates a discrepancy between the report and the BWC footage produces Brady material: the footage that contradicts the report must be disclosed and addressed.
• Giglio v. United States (1972) requires disclosure of impeachment evidence including evidence bearing on officer credibility. An officer's pattern of adopting AI drafts without adequate verification is Giglio material in future cases.
• Chain of custody for documentary evidence now includes the production chain for AI-assisted reports: footage capture, AI processing, draft generation, officer verification, corrections, and submission. A documented chain is defensible. A gap in the chain is a cross-examination opportunity.
• CJIS (Criminal Justice Information Services) Security Policy governs the handling of criminal justice information. AI tools that process CJI must be agency-approved and CJIS-compliant. Consumer tools and personal accounts must not be used with case materials.
• CAD (computer-aided dispatch) and RMS (records management system) are the systems where AI-touched output lives and propagates. Errors that enter these systems reach prosecutors, defense attorneys, courts, and public records requesters. The quality control standard must be met before submission, not after propagation.
• BWC (body-worn camera) footage is the ground truth for AI-assisted report verification. Every factual claim in an AI draft must be traceable to the footage. A claim that cannot be traced to the footage is an unverified claim and must not appear in a sworn report.
• Redaction is a two-failure-mode discipline: over-redaction (withholding too much) produces open-records complaints; under-redaction (missing protected content) produces privacy violations. AI-assisted redaction requires human verification of the output before release.