Detective Yates has twelve hours of interview audio across four recorded sessions. He has a surveillance video log that nobody has finished reviewing. He has three months of phone records, a financial document package, and a case file that has grown to 340 pages. The trial date is in eight weeks. The question he is asking himself at 9 p.m. on a Wednesday is not "do I have enough evidence?" It is "can I get through all of it in time to know what I have?"

This is the investigative load problem that AI is being asked to address in 2026. Not replacing the detective's judgment. Not deciding what the evidence means. Doing the work of getting the detective to the evidence faster, so the judgment can be applied to what the record actually contains rather than what the detective has had time to read.

This lesson covers AI in investigations and analysis: summarization of interview recordings and case materials, pattern extraction from large document sets, and the verification discipline that keeps a fabricated quote or an invented summary finding from sinking a prosecution. The lesson also introduces Brady v. Maryland and Giglio v. United States as the constitutional frame for why getting AI-assisted investigative analysis wrong is not just an embarrassment. It is a disclosure failure.

Terms used throughout: BWC is body-worn camera. RMS is the records management system. CAD is computer-aided dispatch. Brady v. Maryland (1963) is the Supreme Court case requiring prosecutors to disclose exculpatory evidence to the defense. Giglio v. United States (1972) is the case requiring disclosure of impeachment evidence about government witnesses, including officers. Chain of custody is the documented, unbroken record of who had access to and control of evidence at every step from collection to courtroom.

The Investigative Volume Problem

Modern investigations generate enormous quantities of material. A serious criminal investigation in 2026 routinely involves: hours of recorded witness and suspect interviews, BWC footage from multiple officers across multiple contacts with the subject, digital evidence from phones (call logs, messages, photos, app data), financial records, surveillance footage from private and public cameras, social media records obtained through legal process, and a growing pile of investigative reports, supplements, and lab results. A homicide investigation can produce a case file that contains more words than a long novel and more video than a feature film, on a timeline that is compressed by the trial calendar and the rights of the accused to a speedy resolution.

The investigator tasked with this material faces a genuine capacity problem. The highest-value use of an experienced detective's time is analysis and decision-making: recognizing significant contradictions in witness accounts, identifying the lead that opens the next avenue of investigation, developing the charging theory that the evidence supports. The lowest-value use of that same time is mechanical review: transcribing an audio recording, reading a phone record to find the call that preceded the event, re-reading a report for the third time to find a date. AI can do the mechanical review faster and with greater consistency. The investigator's job is to apply professional judgment to what the AI found.

That division of labor is legitimate and useful if it is done with verification discipline. It becomes dangerous when the AI's summary substitutes for the investigator's firsthand engagement with the source material, because the AI, as earlier lessons have established, can generate a plausible summary of content it has not accurately extracted, and the difference between a summary drawn from the record and a summary that fills gaps with inference is not visible on the surface of the text.

Interview Summarization: The Mechanics and the Limits

Audio interview summarization is the AI application with the clearest direct value in the investigative context. An AI tool that can produce a structured summary of a two-hour recorded interview in minutes, identifying the key statements, the timeline the subject provided, the inconsistencies in their account, and the points where the interview turned is genuinely useful. It gets the investigator to the substance of the interview faster and with a structural overview that aids analysis.

The mechanics are similar to report-writing summarization: the audio is transcribed, the transcript is processed by a language model, and the model produces a structured summary with, in more sophisticated implementations, timestamps pointing to the underlying audio segment for each summarized claim. The summary is not the record. The transcript and the audio are the record. The summary is a navigation tool that helps the investigator engage with the record more efficiently.

The Quote Fabrication Risk in Interview Summaries

The most serious failure mode in interview summarization is the fabricated or distorted quote. A language model summarizing an interview will, if not constrained, sometimes produce direct quotations that are composites of what the subject said, paraphrases presented in quotation marks, or in the worst case, invented statements that fit the narrative of the interview but were never spoken. This failure mode is not an edge case. It is a documented behavior of language models when asked to summarize recorded content, particularly when the audio was unclear, when the summary prompt asked for the "key statements" from the interview, or when the relevant statement was a paraphrase rather than a verbatim exchange.

The legal consequence of a fabricated quote in an investigative summary is severe. If a case summary or investigative report attributes a statement to a suspect or a witness that the person never actually made, and that statement is used to support charging, or to brief a prosecutor, or to influence a warrant application, the downstream consequences can include: suppression of evidence that was sought or obtained based on the incorrect summary, Brady disclosure obligations to provide the actual interview alongside the summary, and Giglio obligations if the witness whose statement was distorted is a government witness who will testify. A fabricated quote that reaches a charging document or a sworn affidavit is not a quality problem. It is a constitutional problem.

Every quote in an AI-generated investigative summary must be verified against the audio and the transcript before it is used in any legal document or action. There is no expedient exception.

The verification standard for quotes in investigative summaries is absolute: no AI-generated quote should appear in any investigative report, case summary, charging document, warrant affidavit, or prosecutor briefing unless a human has confirmed that the exact words appear in the audio recording and the transcript. If the AI produces a summary that includes quoted language, the investigator must locate the timestamp in the recording, play the audio, confirm the words, and confirm the context before that quote enters the case file as a representation of what the subject or witness said.

Inconsistency Detection: Where AI Adds Real Value

One of the more genuinely useful applications of AI interview summarization is inconsistency detection across multiple interviews or across multiple statements by the same subject. An investigator managing 12 hours of interview audio across 4 sessions with different parties may not have time to cross-reference every statement from Session 1 against every statement from Session 3. An AI tool that produces summaries of each session and can then compare them, identifying where the accounts of the same event differ between sessions or between parties, is accelerating a task the investigator would have to do manually.

The caveat is that AI-identified inconsistencies must be verified against the actual recordings before they are used. The model may identify an "inconsistency" that is in fact a transcription error, a paraphrase of what was actually said, or a difference in emphasis that does not represent a substantive contradiction. An investigator who confronts a witness with an "inconsistency" that turns out to be an AI transcription artifact has undermined the interview and exposed themselves to a credibility challenge. The AI flags inconsistencies for the investigator to evaluate; the investigator confirms them against the source.

Document and Record Summarization for Large Case Files

Beyond audio, AI tools are being used to process and summarize large document sets: phone records, financial records, incident reports, surveillance logs, and multi-volume case files. The operational value is similar to interview summarization: getting the investigator to the significant material faster than manual reading allows.

For phone records in a criminal investigation, AI can process a log of thousands of calls and texts and identify: calls between specific numbers around a specific time window, patterns of contact between identified parties, gaps in contact that may be significant, and calls to or from numbers associated with other parties in the investigation. This kind of pattern extraction from large structured datasets is an area where AI performs reliably because it is closer to database querying than language generation. The failure risk is lower for this application than for audio summarization because the source data is structured and the model is identifying patterns rather than generating narrative.

For unstructured document sets, including narrative reports, the failure modes from report summarization apply. An AI tool that reads 340 pages of case file and produces a 15-page summary may present the summary in language that is more organized and coherent than the underlying material. Apparent patterns in the summary may not be patterns in the evidence. Connections that seem clear in the summary may rest on a detail the model inferred rather than extracted. The investigator must read the summary with a skeptic's orientation: this is what the model found; is it what the record shows?

The verification discipline for document summarization is: any significant finding, any factual claim used in an investigative action, and any conclusion presented to a prosecutor must be traceable to the specific page, document, and record in the source material. A finding that "the subject was in contact with Individual X on the evening of the incident" must point to a specific call log entry, a specific text message, or a specific record. "The AI summary said so" is not an adequate investigative citation.

Brady, Giglio, and Why AI Summarization Is a Constitutional Matter

The constitutional frame for AI in investigations is Brady v. Maryland and Giglio v. United States. These are not abstract legal doctrines. They are the operating rules for every investigation and every prosecution in the country, and AI in the investigative workflow directly implicates them.

Brady v. Maryland (1963) established that the prosecution must disclose material exculpatory evidence to the defense. Material exculpatory evidence is any evidence that, had it been disclosed, might have changed the outcome of the proceeding. The Brady rule is not limited to documents in a file folder. It extends to all evidence in the government's possession or control, including evidence known to the investigating agency. If AI summarization causes an investigator to miss exculpatory evidence in a case file because the model's summary did not surface it, the Brady obligation has not been met. The fact that the AI did not include it in the summary is not a defense; the obligation runs to what was in the record, not what the AI reported.

Giglio v. United States (1972) extended Brady to cover impeachment evidence about government witnesses. If a government witness has made inconsistent statements, has a deal with the prosecution, or has credibility problems, those facts must be disclosed to the defense. In the AI context, Giglio creates a specific risk: if an AI interview summary distorts or omits the inconsistencies in a government witness's account, the investigator may not flag those inconsistencies to the prosecutor, and the prosecutor may not disclose them to the defense. The chain from AI summary error to Brady-Giglio disclosure failure is short and direct.

There is also a disclosure question about the AI tools themselves. The King County (Washington State) prosecutor's policy on AI-written police reports addressed this directly: when AI touches the evidence, the defense may be entitled to know how it was used and what was reviewed. In the investigative context, this question is still being worked out in courts and prosecutorial guidance, but the direction of travel is clear: transparency about AI involvement in producing the summaries and analyses on which charging decisions rest is coming, and agencies that have built verification and documentation practices into their AI workflows will be better positioned to answer those questions than agencies that have not.

The Investigative Audit Trail

An investigative audit trail, the documented record of how AI was used, what it produced, how the output was verified, and what human decisions were made on the basis of the AI's findings, is not optional if an agency wants to use AI in investigations defensibly. It is the mechanism that allows the investigator to answer the question "how did you reach this conclusion?" with a documented, checkable account rather than a reference to a summary document no one can trace to the underlying evidence.

What should the audit trail capture? At a minimum: which AI tool was used for which purpose (audio summarization, document analysis, pattern extraction), the date and input materials provided to the AI, the output produced (the summary or analysis document), the verification steps performed by the investigator (which quotes were confirmed against the audio, which findings were traced to source documents), and the investigative decisions made on the basis of the verified findings. This documentation serves two purposes simultaneously: it is the investigator's protection if the AI produced an error they caught, and it is the agency's exposure if the AI produced an error they did not catch and the verification protocol was not followed.

The audit trail principle connects back to the broader theme of this chapter and this program: AI in public safety is not primarily a technology story. It is an accountability story. The technology is the tool. The accountability structure, the verification, the documentation, the human decisions at every step, is what determines whether the tool produces better outcomes or just faster ones. In the investigative context, faster outcomes that rest on unverified AI summaries are not better outcomes. They are faster paths to disclosure failures, wrongful charges, and collapsed prosecutions. The goal is better outcomes: investigations where AI got the detective to the evidence faster, the detective verified what AI found, and the case that went to trial was built on the actual record, not the model's version of it.

Key Takeaways

• AI in investigations addresses the volume problem: getting investigators to the significant material in large case files faster, through audio interview summarization, document analysis, and pattern extraction from structured records. The goal is to free investigative judgment for analysis, not to replace it with automated findings.
• The fabricated or distorted quote is the most dangerous failure mode in AI interview summarization. Every quote in an AI-generated investigative summary must be verified against the audio recording and transcript before it appears in any legal document, charging paper, or prosecutor briefing. There is no expedient exception.
• Brady v. Maryland requires disclosure of material exculpatory evidence. If an AI summary fails to surface exculpatory evidence that was in the case file, the Brady obligation has not been met. The fact that the AI did not include it is not a defense; the duty runs to what was in the record.
• Giglio v. United States requires disclosure of impeachment evidence about government witnesses. AI summarization errors that omit or distort inconsistencies in witness accounts can create Giglio disclosure failures when those witnesses testify.
• Any significant investigative finding, any factual claim used to support a charging decision or warrant, and any conclusion presented to a prosecutor must be traceable to the specific page, document, or record in the source material. "The AI summary said so" is not an adequate investigative citation.
• An investigative audit trail that documents AI tool use, inputs, outputs, verification steps, and human decisions is the mechanism that allows the investigator to answer "how did you reach this conclusion?" with a documented, checkable account. It protects the investigator when the verification was done correctly and exposes the gap when it was not.
• The King County prosecutorial standard and the direction of Brady-Giglio disclosure doctrine both point toward increasing transparency requirements about AI involvement in investigative and evidentiary work. Agencies that build verification and documentation practices into their AI workflows now are building a defensible foundation for the accountability questions that are coming.