Keeping interpretation accountable when using AI

AI can help a researcher organize interview material and compare possible themes. I would give it tasks with inputs and outputs we can check, then return to the source material before accepting an interpretation.

Qualitative analysis involves deciding what an account means in context. We need to understand the question a participant was answering and the conditions around the experience. A concise summary can help us find our way through the material, provided we can check what was included and what was left out.

Identify the work that needs support

Before adding AI to a workflow, I would ask where the researcher is spending time. Are they locating excerpts, comparing accounts, checking a theme, or preparing a report? Each task calls for a different kind of support.

If the problem is finding relevant passages, a longer generated summary may add work. I would focus the tool on retrieval and ask it to preserve the references needed to review each passage. If the team is comparing possible themes, I would ask for tentative groupings and the material supporting each one.

Give the tool a bounded task

Useful tasks to explore include indexing passages, comparing sessions, proposing preliminary clusters, and locating accounts that challenge a developing interpretation. The researcher needs a way to inspect the source material behind the output.

I would keep a record of:

• the task and material supplied to the tool
• the output it produced
• the passages checked during review
• the researcher’s changes and reasons for accepting or rejecting a suggestion

That record makes the division of work clear. It also helps another researcher understand how the interpretation developed.

Keep the source close

A useful theme should lead you back to the material that supports it.

Check the evidence behind each claim

A summary may make a pattern sound more consistent than the interviews support. I would check whether a proposed theme rests on one detailed account or appears across several sessions. Similar language can also mean different things in different circumstances.

Keeping references attached makes those checks possible. It gives the researcher a way to examine qualifications that may have disappeared in the shorter version and to explain why the final wording fits the evidence.

Look closely at disagreement

Repeated ideas deserve attention, and so do accounts that differ. Someone who completed a task because they had unusual access may help explain a condition that other participants never mentioned.

During review, I would look for:

• accounts that challenge the leading interpretation
• similar outcomes reached for different reasons
• changes within the same participant’s explanation
• evidence that should lower confidence in a claim

The goal is to understand where a finding applies and where it may break down. A disagreement can suggest a follow-up question or show that the theme needs narrower wording.

Keep responsibility for interpretation clear

The researcher remains responsible for the research question and for judging whether the material can support the conclusion. That includes how uncertainty is reported and what the team recommends as a result.

I would use a final review to compare the proposed account with the original material, check the limits of each claim, and record unresolved questions. The people using the findings should be able to trace a recommendation through that record.