An AI transcript is not the research data by itself. It is a derived file whose reliability depends on the source audio, consent process, transcription convention, corrections and version history.
This guide focuses on provenance: being able to explain where every quotation, code and theme came from.
The research provenance chain
| Stage | Required record | Main control |
|---|---|---|
| Recruitment | Participant reference and approved contact route | Collect only necessary identifiers |
| Consent | Current participant information and consent record | Explain recording, processing, reuse and withdrawal accurately |
| Capture | Original audio plus interview metadata | Approved device, secure transfer and file integrity |
| Transcription | Draft transcript and service details | Defined convention and source-audio checking |
| Pseudonymisation | Analysis copy and separate identity key | Consistent replacements and restricted access |
| Analysis | Codebook, coded excerpts and analytic memos | Trace every claim back to source passages |
| Publication | Checked quotations and disclosure-risk review | Context, accuracy and deductive-identification check |
| Retention | Disposition log for every data type | Apply the approved data-management plan |
Ethical agreement and data-protection basis are different decisions
A participant may agree ethically to take part and be recorded while the research organisation relies on public task or legitimate interests as its UK GDPR lawful basis. The ICO’s current research guidance explains that research processing still needs a lawful basis and appropriate safeguards.
Document separately:
- why participation and recording are ethically acceptable
- the lawful basis for personal-data processing
- the condition for special-category or criminal-offence information
- what withdrawal means at each research stage
- whether future reuse or archiving is included
See the ICO’s current research provisions guidance.
The AI service is part of the method
Replacing approved local transcription with a cloud AI service can change the processor, data location, subprocessors, retention, model-training use and withdrawal process. That change may require institutional, sponsor or ethics review.
Before uploading an interview, confirm:
- the service is approved for the project
- a suitable data-processing agreement exists
- processing and storage locations are known
- customer content is not reused beyond the approved purpose
- audio, transcript and backups can be deleted as required
- files can be exported in usable formats
- workspace access and support access are controlled
Choose the transcription convention before correcting text
A transcript is shaped by decisions about pauses, false starts, dialect, grammar, laughter, overlap and non-verbal events. Choose a convention that matches the research question:
- Verbatim: preserves repetitions and speech patterns.
- Intelligent verbatim: removes some disfluency while preserving meaning.
- Conversation-focused: records overlap, pauses and interaction detail.
- Content-focused: prioritises substantive meaning over delivery.
Apply the same rules across the dataset and record any later change.
The transcript-verification pass
- Keep the original audio unchanged.
- Create the machine transcript as a clearly labelled draft.
- Listen while checking every section required by the method.
- Correct names, numbers, technical terms, negations and speakers.
- Mark unintelligible speech consistently rather than guessing.
- Add contextual field notes where audio is insufficient.
- Save a version used for analysis so later edits do not silently alter the evidence.
The UK Data Service provides practical guidance on research transcription.
Pseudonymisation is not anonymisation
Replacing a participant’s name with “P07” does not make the transcript anonymous when an identity key exists or the person can be recognised from occupation, location or rare events.
A pseudonymised transcript remains personal data. Keep the identity key separately, restrict access and review quotations for deductive disclosure before publication.
The quotation card
Every publishable qualitative quotation should have:
- participant code
- interview and timestamp reference
- verified wording
- relevant question and surrounding context
- anonymisation changes
- analytic code or theme
- reason for selection
- disclosure-risk check
This prevents a striking phrase from becoming detached from the participant’s intended meaning.
AI-assisted coding: require source links and negative cases
AI can suggest codes, group passages and summarise patterns. Its output should remain exploratory until the researcher checks:
- which excerpts support the proposed theme
- which excerpts contradict it
- whether minority perspectives were flattened
- whether the sample supports the conclusion
- how the model and prompt affected the output
- whether the approved protocol permits that processing
An AI-generated theme is not a finding. The audit trail should show how the researcher accepted, changed or rejected it.
Focus groups need an additional speaker map
Use participant codes, a seating plan and moderator notes. Overlapping speech and unequal microphone distance make automatic speaker labels unreliable. Check attribution manually before analysing a statement.
Also explain that the researcher cannot guarantee that other participants will maintain confidentiality outside the group.
Retention by data type
Do not apply one deletion date to everything automatically. The approved plan may distinguish:
- original audio
- machine transcript
- verified transcript
- identity key
- analysis dataset
- codebook and memos
- publication quotations
- archival or shared version
Keep each item only for its justified research, verification, funder or archival purpose.
Where NERALVO Halo fits
NERALVO Halo is a slim magnetic recorder with 64GB local storage, NOTE and CALL modes, up to 35 hours of recording and Bluetooth transfer to DOWAY.
DOWAY can create transcripts, summaries, speaker-separated notes, templates, translations and mind maps. The current package includes one year of DOWAY Max from activation.
NOTE mode may support approved interviews, focus groups, research meetings and private field reflections. CALL mode is supported rather than universal. Local capture does not make later DOWAY processing local, so the complete route must be included in research approval and participant information.
Review the current NERALVO Halo details.
Bottom line
An AI recorder can reduce transcription effort without weakening qualitative research only when every derived file remains traceable to the approved source and method.
Preserve provenance, define the transcription convention, distinguish pseudonymisation from anonymity, verify quotations and document every AI-assisted analytic step.
Ready to capture meetings properly?
View the NERALVO Halo AI voice recorder with 64GB local storage, meeting capture, compatible phone-call recording workflows and one year of DOWAY Max included.
View NERALVO Halo