What the Researcher Still Must Know
When a qualitative researcher opens a transcript and asks a generative AI system to identify themes, the result may look like a finished analysis. The system can group passages, provide headings, and write a summary in a tone familiar from published research. The difficult question begins afterward: what must the researcher still know to decide whether any of those themes deserves to appear in a paper?
A recent survey offers a revealing starting point. In an October 22, 2026 article for Social Science Space, Michael Todd reported on Lumivero’s survey of 115 researchers. The sample was relatively small and recruited for a commercial research-software company’s study, so its percentages should not be mistaken for a representative picture of a discipline. Within that sample, however, a clear tension appeared.
Respondents reported using AI for certain tasks while expressing much less confidence in its role in interpretation and sense-making.
Todd highlighted an especially useful distinction. Researchers were more comfortable where they could verify a system’s result against accessible evidence and less comfortable when the task required interpretive judgment. This pattern deserves attention because it suggests a better question for institutions than how frequently their researchers use new tools. We should ask whether researchers retain the skills to detect, explain, and correct the tool’s analytic errors.
Consider a hypothetical interview study of why employees decline voluntary training. An AI system reads 12 anonymized transcripts and proposes a theme called “resistance to change.” It points to statements about scheduling conflicts, a supervisor’s inconsistent support, unclear benefits, and fear of embarrassment. The summary may appear coherent. But those statements could point to four different social processes. Some describe constraints on time, some institutional incentives, some doubts about relevance, and some a threat to professional identity.
The danger lies in how quickly a plausible label can become the starting point for the rest of the work. If the researcher accepts “resistance to change,” later summaries may reproduce it. The codebook may harden around it. A report might recommend motivational communication when the strongest evidence actually supports changing the schedule or clarifying a manager’s obligations. No fictional case is needed to establish that this is an important possibility: it follows from what happens when interpretive categories direct attention.
The appropriate skill is the ability to contest a category at its point of formation. A researcher should be able to ask which passages are grouped together, which cases resist the grouping, and what alternative account would explain the same material. These questions are not ceremonial human oversight. They are the work of analysis.
Research on cognitive offloading helps clarify the challenge. In a 2016 review in Trends in Cognitive Sciences, Evan Risko and Sam Gilbert described how people move mental work into external actions or devices. Offloading can be adaptive, and the authors examined how people’s assessments of their own abilities shape those choices. A notebook, calendar, transcription application, or digital search tool can make inquiry more effective. The institutional problem emerges when a person loses the capacity to judge the external support without noticing that the capacity has weakened.
Generative AI adds a distinctive feature to this familiar arrangement: an external aid can now supply a polished explanation, not simply a stored reminder or retrieved document. Its grammar and organization can confer an appearance of confidence. The researcher may be tempted to treat a coherent synthesis as if it were a validated inference.
Fluency makes an answer easy to read. It cannot establish that the source evidence supports the answer.
One response is to ban assistance from interpretive work altogether.
Some qualitative scholars have made strong objections to generative AI in reflexive analysis, and Todd’s report discussed an open letter published in Qualitative Inquiry. Their concerns about situated meaning, social responsibility, and the role of the analyst deserve serious engagement. Yet a blanket ban can also prevent inquiry into whether a tool might help researchers locate overlooked counterexamples, compare rival codebooks, or document disagreement more systematically.
The more useful distinction concerns the location of interpretive authority. A research team can use a system to propose three competing names for a pattern while retaining responsibility for deciding whether the pattern exists. The team can invite a machine to flag passages that contradict its draft codebook, then inspect those passages independently. It can also keep a record of which suggestions survived confrontation with source material and which were discarded.
In each case the tool supplies a candidate; the researcher must provide the justification.
This change has consequences for how teams train junior scholars.
Imagine a novice analyst who has only ever encountered a machine’s completed theme table. They may know how to request better output, yet struggle to say why an apparently persuasive category is sociologically thin. That is a problem of apprenticeship. A field that automates its easiest interpretive exercises before novices practice them risks producing competent operators who lack an independent standard of quality.
The remedy begins with deliberate contrast. Give a novice a modest set of interview passages and ask them to code it without assistance. Have them write a short memo explaining why a particular excerpt fits a theme and what would make them reconsider. Only then introduce a tool-generated interpretation. Compare the two accounts, identify discrepancies, and ask the novice to defend a final decision from the transcript. The point is to create a learning opportunity that remains available even when production work is assisted.
Institutions should also make this competence observable. Counting the number of transcripts processed tells managers little about the quality of interpretation. A stronger evaluation asks whether analysts can trace a finding to source passages, identify an important negative case, distinguish description from explanation, and articulate how a theme changed under challenge. These are tasks a trained reviewer can inspect, and they fit the norms of rigorous research without pretending that one metric captures all quality.
Research involving human participants adds another layer. The ethical consequences of an interpretation depend on context, consent, and downstream use. A quote about workplace fear may carry a different meaning when detached from the speaker’s relationship to an employer.
A system can assist in locating similar expressions, but neither a similarity score nor a confident summary resolves whether using that quote is fair to the participant or adequate for the research question.
A well-designed audit trail should therefore record more than which software produced an output. It should show when the researcher accepted, changed, or rejected interpretive suggestions and why. If a finding later proves weak, collaborators should be able to trace the decision rather than discover that responsibility disappeared into a vague statement that the analysis was AI-assisted. Such records also support training: a student can learn from a documented correction in ways that are impossible when the only retained artifact is a finished report.
The question of productivity still matters. In a 2023 experiment published in Science, Shakked Noy and Whitney Zhang found substantial gains in speed and rated quality for specific professional writing tasks performed with ChatGPT. That result gives organizations reason to explore tools for work with clear evaluative standards. It does not establish equivalent gains in qualitative interpretation, where the central challenge may be to decide what the standard ought to be. We should carry the evidence into new settings without pretending the settings are identical.
There is an institutional incentive to blur the difference. A research office can readily display how many staff members adopted a tool, how many hours they report saving, or how quickly a first draft appeared.
Demonstrating that an analyst detected a mistaken assumption is harder. But the difficult measure is often closer to what research exists to accomplish. A fast report can satisfy an administrative demand while being wrong about the people it describes.
The practical alternative is an apprenticeship in correction.
Institutions should preserve a small portion of unassisted analysis, build structured comparisons with generated interpretations, and reward researchers for documenting errors they caught before publication. Editors and supervisors can ask for the rationale behind a classification, not merely a statement that someone reviewed it.
These habits train the skill that remains useful when a tool’s interface or underlying model changes.
Todd’s report put a name to an observable tension: adoption without confidence. The next phase of research practice should make confidence something scholars can earn through demonstrable judgment. A qualitative researcher does not need to perform every routine step unaided. They do need to know when the beautiful summary is a bad analysis, how to prove it, and what to do next.
