Can I Use AI Transcription for IRB-Approved Research?

What IRBs Actually Evaluate

IRBs don't approve or reject AI transcription as a category. Georgia State University's IRB guidance is explicit: researchers who use AI transcription must describe the platform in their IRB application as a data-sharing practice. Teachers College Columbia University's IRB takes the same position, requiring researchers to proactively assess and disclose any AI-based software that interacts with human subjects data, whether directly or indirectly. NIH's own IRB guidance confirms the logic: no matter how you use AI in your research, you need to clearly explain its use in your study plan so the IRB can properly review it. The review focuses on your plan, not on AI as a category.

"Any AI-based software that interacts with human subjects data, whether directly or indirectly, may require both a security assessment and IRB review prior to use in research."

Teachers College Columbia University IRB Guidelines, 2026

When you submit a protocol that mentions AI transcription, here's what reviewers will want answered, and how Qualtranscribe's answers hold up:

These are the same questions an IRB asks about human transcription services, with one addition specific to AI: whether automated processing introduces any new privacy risk not covered in your original consent language.

The AI Training Problem

This is the one that catches most researchers off guard.

Many AI transcription platforms, including some widely used ones, include language in their terms of service allowing them to use uploaded audio to improve their models. This is legal. It's disclosed, usually in the fine print. And it is directly incompatible with IRB-approved research, because your participants consented to have their recordings used for your study, not to train a commercial AI system.

Some of the most popular general-purpose transcription platforms grant themselves perpetual rights to use uploaded content to train their proprietary speech recognition models by default, with opt-out requiring a direct email to their support team. This is not a data breach. It's a disclosed policy. But it is the kind of policy that conflicts with a research consent form that promises confidential handling of participant data.

The fix isn't to avoid AI transcription entirely. It's to use a platform with a documented no-training policy. Qualtranscribe does not use recordings for AI training on any plan, including Free. That's a verifiable, citable fact for your data management plan rather than something buried in terms-of-service.

What Your Data Management Plan Needs to Say

If you're including AI transcription in an IRB protocol, the data management section needs to address these points specifically. Here's sample language you can adapt:

Sample IRB Data Management Language

"Interviews will be transcribed using Qualtranscribe's Instant Draft AI transcription service. Audio files are processed on Qualtranscribe's HIPAA-compliant infrastructure and are not used to train AI models on any plan. Files are stored in [us-east-1 Northern Virginia / eu-central-2 Frankfurt / ap-northeast-1 Tokyo] and will be deleted within 30 days of project completion. A Business Associate Agreement is in place for this study. Transcripts will be reviewed by the research team for accuracy before entering the coding phase. No recordings will be shared beyond the transcription service or used for any purpose other than this study."

The key elements: name the platform, document data residency, confirm the no-training policy, state the retention window, and reference the BAA or data processing agreement. Generic language like "a secure transcription service" won't satisfy most IRB reviewers in 2026.

When AI Transcription Works Well for IRB Research

AI transcription fits IRB-approved research in several specific situations, and where it doesn't is equally worth knowing.

Qualtranscribe's Instant Draft is specifically built for the left column. A transcript appears in minutes, alongside Smart Insights that automatically surface themes, key quotes, and sentiment patterns without manual coding. For researchers managing tight timelines, that changes what's possible before a first team meeting. For sessions where the transcript needs to hold up under formal scrutiny, human transcription from the same platform is available per session, rather than as a platform-level commitment.

The Consent Form Question

One detail that comes up less often than it should: if you're using AI transcription, your consent form should say so.

Not in technical detail. Participants don't need to understand speech recognition models. But your consent language should mention that interviews may be transcribed using automated tools and that any third-party processing is covered by confidentiality agreements. IRBs increasingly expect this transparency, and participants have a reasonable interest in knowing their voice recordings are processed by software rather than only by named research team members.

A workable addition to existing consent language: "Recordings will be transcribed by a secure, HIPAA-compliant service under confidentiality agreement. Recordings are not shared beyond the research team and are not used for any purpose other than this study."

What This Looks Like in Practice

A qualitative researcher running 25 interviews for a funded NIH study uploads each recording to Qualtranscribe's Instant Draft immediately after each session. Transcripts come back in minutes, with Smart Insights flagging recurring themes. The research team reviews each transcript against the recording before it enters the coding phase. The data management plan submitted to the IRB names Qualtranscribe, documents the no-AI-training policy, confirms HIPAA compliance and the BAA, and specifies 30-day file deletion. The IRB reviews this as a standard data-sharing disclosure rather than an exceptional case requiring additional scrutiny.

That's what IRB-compatible AI transcription looks like when it's done right. Ready to build that workflow? Get started here and the free plan includes 75 minutes to test it before committing.

FAQ

Do I need to amend my IRB protocol if I switch from human to AI transcription mid-study? Usually yes. Adding a new third-party processor of identifiable data is a change to your data management plan, which typically requires an amendment. Check with your IRB office before switching.

Can AI transcription be used for HIPAA-covered research? Yes, if the platform holds a signed Business Associate Agreement and meets HIPAA requirements. Qualtranscribe is HIPAA compliant and provides BAAs for health research. The BAA is what makes the workflow legally compliant, not just the platform's general security practices.

What if my IRB says no to AI transcription? Ask what the specific concern is. Most IRB objections come down to one of three things: the AI training issue, data residency, or consent language. All three are addressable with the right platform and protocol language. A blanket no-AI-transcription policy is worth pushing back on with specific documentation.

Is AI transcription less accurate than human transcription for research? Accuracy depends on audio quality, speaker count, accent, and technical vocabulary. On clear audio with one or two speakers, AI transcription from a research-grade platform performs well. On focus groups with overlapping speech or recordings with heavy background noise, human transcription is more reliable.

Does using AI transcription affect the methodological rigor of my study? Only if the output is used without verification. AI transcripts should be reviewed before entering the coding phase, the same way a human transcript would be spot-checked against the recording. The methodological commitment is to accuracy, not to any specific transcription method.

Related Reading

• The Ethics of AI in Academic Research Transcription

• Understanding Artificial Intelligence Through the IRB Lens

• IRB Transcription Mistakes Researchers Make

• HIPAA-Compliant Transcription for Research

• De-identification, Anonymization, and Pseudonymization: What's the Difference