Source
urlhttps://doi.org/10.1177/10497323251389800
rawraw/dellafiore-et-al-2025-artificial-intelligence-in-qualitative-research-insights-from-experts-via-reflexive-thematic.pdf

TL;DR: Fourteen expert qualitative researchers from Italian academic and healthcare settings were interviewed about AI. The study surfaces a culture of concealment around AI use, introduces the “illusion of meaning” as a practitioner concept, and shows that the human–AI divide maps cleanly onto a technical tasks / interpretive tasks split that experts negotiate with ambivalence rather than resolve.

Problem

Most research on AI in qualitative research either tests AI tools against human outputs or surveys researchers’ perceptions quantitatively. What is missing is an in-depth, qualitative account of how expert researchers — people with ten or more years of qualitative practice, doctoral training, and active publication records — actually experience, interpret, and negotiate AI integration within their methodological commitments. This study fills that gap using the most appropriate tool for the job: semi-structured interviews and reflexive thematic analysis applied to the phenomenon itself.

The setting matters. Italian socio-anthropological and healthcare contexts are underexplored in a literature dominated by Anglophone institutions. Participants here include anthropologists, sociologists, physicians, psychologists, and nurses — a disciplinary range that captures real methodological diversity. The healthcare-qualitative crossover context also surfaces concerns about environmental sustainability and data governance that appear more prominently than in methodological debates within social science alone.

Approach

Semi-structured interviews with 14 expert qualitative researchers, conducted April–May 2025 (face-to-face or Microsoft Teams). Purposive sampling targeted researchers with at least two to three completed qualitative projects, peer-reviewed publication records, and active qualitative responsibilities. Mean age 46 years; 7 men, 7 women; near-universal PhD level; all with 10+ years of research experience. Nine specialized in socio-anthropological fields; five in medicine and healthcare.

Notably, “Artificial Intelligence” was deliberately left undefined at the start of interviews — a methodological choice designed to elicit participants’ own conceptualizations rather than constrain responses. The opening question was simply: “What is being said about AI in your qualitative research environment?” Real-time reflective probing clarified what participants meant without imposing definitions.

Data were analyzed using Braun and Clarke’s six-step RTA framework, following RTARGs (Reflexive Thematic Analysis Reporting Guidelines). Two researchers coded independently; discrepancies resolved through discussion; senior researchers consulted for coherence. Member checking with three participants confirmed the thematic structure required no significant amendment.

Four themes emerged:

  1. Perspectives on AI and Qualitative Research — ambivalent, polarized attitudes including technophobia, prejudice, and openness; a pervasive shame around declaring AI use; skill-related barriers; and a pragmatic call for control
  2. Anthropological and Philosophical Dimension — the human–machine distinction (capacities vs. skills); AI as “alter ego” or “additional voice”; risks to researcher identity; the “illusion of meaning”
  3. Pragmatic Dimension of AI Usage — AI as support for technical tasks (transcription, translation, writing); contested role in interpretive phases; limits on fieldwork and ethnographic research
  4. Ethical and Sustainability Dimension — Western epistemological dominance, data ownership and governance gaps, ecological costs of AI infrastructure

The “Illusion of Meaning”

The most conceptually distinctive contribution is a term that emerged from participants, not from the researchers’ framework: the “illusion of meaning.” One participant articulated it precisely: “AI could, in some way, create the illusion of certain meaning aggregates — it might reconstruct meaning in an interview from the interviewee’s perspective. And that would be something new. I call it an ‘illusion of meaning’ because, clearly, these are algorithmic inferences and combinations.”

This is a practitioner-level articulation of what epistemic-flattening describes at the architectural level. The concern is not that AI is obviously wrong, but that it is subtly misleading — producing outputs that look interpretively meaningful but lack the interpretive grounding that comes from researcher immersion. The participant who coined the term described it with “a firm tone of voice and intense eye contact” — field note evidence of the depth of concern. The illusion is precisely that: it is not obviously wrong, which makes it harder to detect and resist than a hallucination would be.

The Shame Culture Around AI Use

A striking empirical finding is the culture of concealment around AI use. Seven of fourteen participants acknowledged widespread prejudice against AI in their professional circles. Several admitted to a “fear of being judged” that led them to conceal AI use — “in the darkness of our little rooms.” More revealingly: 13 of 14 participants admitted using AI in some form, but several had initially presented themselves as non-users, only clarifying during the interview that they were already employing AI-based tools. One realized mid-interview that NVivo (now Lumivero) itself uses AI.

This has methodological implications: research surveys asking about AI use may systematically underreport it. The Paulus & Marone (paulus-marone-qdas-discourse-2024) finding — that researchers may be absorbing AI epistemologies through platforms without recognizing it — gains empirical texture here. Researchers who don’t know they’re using AI cannot critically evaluate what it’s doing to their analysis.

Technical Tasks vs. Interpretive Tasks

The clearest consensus was a functional division of AI labor:

  • Near-universal acceptance: transcription (all 14), scientific English writing and Italian-to-English translation (13/14), document summarization, visual abstracts
  • Contested: study design, literature review (hallucination and source fabrication concerns), coding and theme generation
  • Strong resistance: interview analysis described as a “black box” by several; ethnographic fieldwork explicitly excluded (“AI can’t do that on its own, because it doesn’t have a body”)

This maps onto the small-q / Big-Q distinction (nicmanis-spurrier-ai-guide-2025) but arrives there through practitioner experience rather than methodological argument. The phrase “it doesn’t have a body” — offered with a smile — captures the phenomenological commitment to fieldwork immersion that computational tools cannot replicate.

The generational divide identified by chatzichristos-ai-positivism-2025 reappears: younger researchers embraced AI more readily, while senior researchers expressed concern about their capacity to guide younger colleagues in developing critical judgment.

Epistemological Stance

Interpretivist and constructionist — consistent with the Braun & Clarke approach the team applies. The paper does not use intercoder reliability metrics, following RTA’s epistemological commitments. Researcher positionality is acknowledged. AI appears as an object of study, not a tool used in the analysis.

The ethical theme surfaces a concern with Western epistemological dominance that connects to structural arguments in the corpus: AI’s training data reflects “prevailing epistemologies, which are predominantly Western,” risking the marginalization of “minority viewpoints from subcultures or underrepresented groups.” This connects directly to epistemic-flattening, dahal-genai-qualitative-nepal-2024, and sakaguchi-chatgpt-japanese-2025, but is articulated here from within a practitioner community — not as an external critique but as a live ethical concern.

Rigor and Trustworthiness

The study follows RTARGs, combines dual independent coding with senior researcher oversight, integrates field notes systematically, conducts member checking, obtains ethical approval (Regional Ethics Committee of Umbria), and complies with GDPR. Consistent with RTA epistemology, quantitative reliability metrics are not reported — explicitly and appropriately.

The limitation acknowledged most honestly: this is a study of perceptions, not practices. The researchers cannot observe what participants actually do with AI; they can only analyze what participants report and how they describe it. The open definition of “AI” was a productive methodological choice, but it means participants may have been describing different tools under the same label.

Limitations

  • Geographic and disciplinary scope: All participants from Central and Northern Italy, predominantly healthcare-adjacent disciplines. Transferability to other national or disciplinary contexts is limited.
  • Perception–practice gap: Self-reported use is not observed use. The shame culture the paper itself identifies means participants may have underreported AI use even in this relatively safe research context.
  • Undefined “AI”: The deliberate non-definition enabled rich conceptual exploration but limits comparability across participants (ChatGPT and NVivo/Lumivero are not the same kind of tool).
  • Rapid obsolescence: Perceptions captured April–May 2025 may already be outdated; the paper notes this explicitly.
  • No comparative data: No benchmarks against human-AI output differences; this is entirely the qualitative perception literature, not the empirical performance literature.

Connections

What links here