Can you use AI transcription for research interviews?

Yes. AI transcription is a sound way to produce the first draft of a research interview. Two conditions decide whether it is acceptable for your study. You must check the draft against the recording before you quote or code any of it, and you must settle consent and data protection before the audio goes to any service.

A checked AI draft is a sound starting point, not a final record

Automatic tools do best on clean, single-speaker audio, and they do worse on conversation. On the CallHome test set, which records conversations between friends and family, Whisper Large V2 had a word error rate of 17.6 percent ¹. Word error rate compares a tool's output with a checked transcript, so a figure like this tells you how many words will need attention on that kind of audio.

Review is not optional for research. Wollin-Giering and colleagues, who tested automatic tools on interview recordings, recommend that you review the text of any transcript to make sure it is accurate, whichever tool produced it ². Spend your checking time where the draft is weakest: names, numbers, negatives, and passages where two people talk at once.

Test the tool on a recording like yours before the study

Northwestern University's research guidance says it is always a good idea to test the service you plan to use with a recording similar to your research data, to see whether its accuracy is good enough ³. Record a one-minute test with the same microphone, in the same room, with a voice like your participants'. Compare the output with the audio, count what you had to fix, and decide whether that effort is acceptable before you record the real interviews.

Settle consent before you press record

The consent rules depend on where the people are. The federal Wiretap Act says that it is not unlawful to intercept a communication when you are a party to it, or when one party has given prior consent ⁴. California is stricter. Its Penal Code 632 punishes anyone who intentionally records a confidential communication without the consent of all parties ⁵. An interview held in California, or with a participant there, should be treated as needing everyone's consent. This is general information, not legal advice. Our guide to recording consent laws covers other places, and the glossary explains all-party consent.

Check where the audio and the transcript will be stored

Most of the risk sits with the tool, not the interview. OpenAI says that abuse monitoring logs for its API are retained for up to 30 days by default ⁶. That is one provider's default for one product, and other providers publish their own periods. Ask each provider how long it keeps audio and text, whether it uses your content to train its models, and where it stores the data.

Consumer defaults can also cause trouble. A federal class action filed in August 2025 alleges that Otter.ai's Otter Notebook service, by default, does not ask meeting attendees for permission to record ⁷. These are allegations, not findings, but the lesson holds: read the default settings before you record.

Universities may also treat transcripts as records. The University of Utah's guidance says that any recorded or transcribed content may be subject to public records requests ⁸. Public records law varies by state and country, so ask your own institution.

It depends: when AI transcription is a poor fit

AI transcription suits most one-to-one interviews, but four cases call for a different route or more checking.

• Restricted data. Some datasets carry rules that forbid particular services. A data use agreement or ethics approval can rule out a tool even when its accuracy is good.
• Crowded or overlapping speech. Focus groups, group interviews and calls with background noise are the hardest audio. They need the most checking, and speaker labels will be least reliable.
• Analysis that depends on delivery. Conversation analysis studies pauses, overlap and hesitation. A machine draft is a poor starting point, and our guide to verbatim and clean verbatim transcription explains the styles.
• Protocols that name a transcriber. Some protocols specify who may handle transcripts. Check that before you choose a route.

What to do, in order

1. Confirm approval. Get ethics or data protection approval for the specific tool, and name the service in your consent form.
2. Test the tool on a one-minute recording that resembles your interviews.
3. Record clean audio, with a microphone near each speaker and one track per person where your recorder allows it.
4. Generate the draft, then check every quote and speaker label against the audio, as set out in our guide to interview transcription.
5. Store and delete on schedule. Keep the original recording until your data plan allows deletion.

If you are preparing the transcripts for coding, read our question on how to prepare transcripts for qualitative analysis. If your source is a podcast or a public talk rather than a research interview, our question on how to transcribe a podcast covers the differences. For the speaker labels, the glossary entry on speaker diarization explains what the tools are guessing.