An AI research assistant helps a researcher work with questions and material: drafting a guide, finding passages across interviews, proposing categories, or summarising a session. Its role differs from an AI moderator speaking to participants, although a research platform may offer both.
The useful question is what the assistant can do reliably with the particular material in front of it. A good result on a short, clear transcript does not establish that it will handle overlapping speech, specialist vocabulary, or contradictory accounts equally well.
Give the task a defined scope
A request to “find insights” leaves substantial room for interpretation. Asking the assistant to locate every passage about changing account ownership, preserve the surrounding context, and identify the source session gives the researcher a more inspectable starting point.
Ballpark’s Coach guidance describes support for analysing research material such as responses and transcripts. Whatever tool is used, check which studies and fields it can actually access before interpreting an answer about “all customers”. Material absent from the tool’s context cannot inform its response.
An assistant can also challenge a draft questionnaire or propose follow-up topics. These are suggestions for researcher review, with recruitment, wording, and method still determined by the study’s purpose.
Check claims and omissions
Generated summaries can introduce unsupported details or turn tentative language into certainty. NIST’s generative AI risk profile identifies confabulation among the risks to manage. In research, an especially important check is whether a quoted passage exists and supports the claim made around it.
Review the source, including audio or video when transcript accuracy matters. Check how many distinct participants support a proposed pattern, whether they were answering comparable questions, and which accounts complicate it. Ten excerpts from one interview remain one participant’s contribution.
Omissions deserve attention too. A polished summary may favour repeated, easily classified comments while overlooking an unusual experience that is central to the decision. Search deliberately for contradictory and missing evidence.
Keep responsibility attached to the work
Record the task, relevant model or tool version where available, input scope, and substantive corrections. This helps colleagues understand which parts were generated and which conclusions the researcher has reviewed.
Use only data the organisation is authorised to process in the chosen system, consistent with participant agreements. When the assistant helps with thematic analysis, describe its contribution accurately rather than implying an entirely manual process. The gain is useful assistance with the work; the final account still needs someone who can explain and defend how it represents the evidence.