Should AI Take Notes in an Internal Investigation Interview?

AI notetakers are becoming increasingly common, raising questions about whether the technology is reliable enough to document witness interviews in internal investigations.

But accuracy is only part of the question.

An Interview Is Not a Meeting

AI notetakers were built for meetings. In a typical meeting, nobody’s credibility is being assessed, there is no adverse party, and no one expects the notes to be compared years later against testimony in a deposition. The goal is usually simple: create a useful record of a conversation that no one expects to contest.

An investigation interview is different. The witness may be frightened, angry, embarrassed, or worried about protecting someone. The investigator is not simply recording a conversation, but rather building a record that may later be challenged.

That difference matters.

Notes and Transcripts Do Different Things

The argument for an AI transcript is obvious: it promises to capture everything. The investigator does not have to worry about writing quickly enough or missing a statement.

But an investigator’s notes are not simply an imperfect transcript. They serve a different purpose, reflecting the investigator’s judgment about which answers mattered, which questions the witness was responding to, where the account changed, and what warranted further inquiry.

The better practice, when circumstances allow, is to have two people in the interview: one to conduct it and one to take notes. That lets the interviewer stay focused on the witness, follow up in real time, and pay attention to how the account develops without also trying to create the record. The notetaker can capture the substance of the interview while the interviewer does the work that requires judgment in the room.

A transcript may capture the words, but it makes none of those judgments. That work still has to happen later, often after the interview is over and without the benefit of what the investigator observed in the room.

There is also a practical question about what happens to the record afterward. A verbatim transcript creates a detailed account of the interview, often stored by a third-party vendor. Before using one, an organization should know where that record goes, who can access it, and how long it is retained. Those questions are easy to overlook when a tool is adopted because it makes ordinary meetings more convenient.

Accuracy Is Not as Simple as It Sounds

AI transcription tools can do more than mishear words; they can sometimes generate words that were never spoken. In a 2024 study of Open AI’s Whisper transcription model, researchers found that approximately 1% of the audio transcriptions contained fabricated phrases or sentences that did not appear in the underlying audio. The researchers also found that hallucinations occurred disproportionately in speech containing longer pauses or periods of silence.

That matters enormously in an investigation. A mistaken adjective, a descriptor that was never used, or the wrong speaker attribution can alter how a statement is understood.

The problem may be especially significant when speech is halting or interrupted. Pauses, incomplete sentences, restarts, and silence are common in serious interviews. A witness describing something humiliating or upsetting may stop mid-thought, search for words, and begin again. Those are exactly the moments when an investigator needs the record to be reliable.

Courts have also raised concerns about possible AI-assisted transcription errors. In a July 2026 memorandum decision, the Indiana Court of Appeals identified errors in a trial transcript, including statements attributed to the wrong speakers and names transcribed incorrectly. The court suggested that generative AI may have been involved, although it did not determine the cause of the errors. In an investigation, mistakes of that kind are not minor formatting issues. Who said what may be central to the finding.

The Tool Also Changes the Interview

Even if the transcript were perfectly accurate, there is another issue: what the tool does to the conversation itself.

A second person taking notes does not affect the interview in the same way an AI notetaker can. A human notetaker is part of the interview team. An AI tool, by contrast, is capturing the conversation word for word, and the witness knows it. That can affect how freely the person speaks. Most witnesses are not necessarily trying to lie. They are thinking about how they will be perceived, what their words might look like later, and how carefully they need to phrase an answer. A recorded or AI-transcribed interview can heighten that instinct.

There may also be consent and privacy issues that do not arise in the same way with a human notetaker. Recording laws vary by state, and the use of third-party AI tools raises additional questions about who is receiving the conversation, where the data is stored, and what happens to it afterward. Those issues should be resolved before the interview, not discovered after the fact.

Where AI Can Help

None of this means AI has no place in an investigation. The better question is whether the work it is doing can be checked.

AI can be useful for building a chronology from documents, organizing exhibits, identifying gaps in a draft report, or helping an investigator see whether an issue has been left unresolved. If the tool gets something wrong, the underlying documents are there to check it against.

An interview is different. The conversation happens once. A second person taking notes hears the same conversation as the interviewer and can clarify or supplement the notes afterward. An AI transcript creates a different problem: if it contains something the witness never said, how would anyone know?

That is what I would ask before using an AI notetaker in an investigation interview. If someone would have to return to the audio to verify the transcript, then the audio is the real record and the transcript is a convenience. If there is no independent way to check the transcript, the technology has not eliminated the risk of an incomplete or inaccurate record. It has simply made the error harder to see.