The State of AI in Qualitative and Mixed Methods Research

Qualitative research has always depended on human judgment—the ability to interpret meaning, weigh context, and see patterns across a body of evidence. In 2026, that judgment is increasingly being paired with AI, but not replaced by it. What researchers are working out, project by project, is not simply where AI can accelerate the work, but something more fundamental: where epistemic authority sits when a machine’s output enters the analysis, and how a researcher keeps hold of it.

The question isn’t theoretical—it shows up directly in the data. In our 2026 survey of 115 researchers across academic and commercial settings, 57% of researchers already use AI in some form, yet confidence has not kept pace with adoption. Across all respondents, 87% are concerned about accuracy and hallucination risk, 86% about research integrity, and 85% about data privacy—concerns that intensify, rather than ease, at the analytical core of the work. Trust is highest where AI’s output is easiest to verify, and lowest wherever a task requires interpretation.

Look closely at what researchers are actually doing in that gap, and a picture emerges.

These are not workarounds or reluctant extra labor. They are the practices through which researchers assert and maintain epistemic authority over their analysis—the same interpretive discipline qualitative research has always demanded, now applied to output no researcher authored.

The problem is that most AI tools give these practices nowhere to live. When verification, translation, and pressure-testing happen outside the software—inside documents, parallel analyses, and ad hoc disclosure decisions—the researcher’s judgment is real but invisible, and the rigor it produces leaves no trace.

AI in research should be built around these practices, not despite them: transparent about how it reached a conclusion, traceable back to source material, and never final until a human confirms it. Designed that way, AI doesn’t compete with the researcher’s authority—it makes the exercise of that authority visible, auditable, and defensible.

In this report, learn where AI has earned researchers’ trust, where it hasn’t, and why adoption and confidence haven’t moved together. You’ll also see what these findings mean for the technology this field needs, and the standard we hold our own work to at Lumivero. Whether you’re a researcher deciding how far to trust AI in your own work or the one shaping the policy that governs it for others, this report will give you insight to guide your next step.

Adoption is already widespread, though far from uniform. 57% of researchers already use AI in some form—21% regularly, 36% occasionally—and another 19% have tried it without ever quite making it a habit.

Adoption climbs steadily with experience rather than concentrating among the newest researchers. Looking at regular use as a share of each experience bracket, researchers with 11–20 years in the field have the highest rate of regular use (33%), followed by those with more than 20 years (30%) and those with 6–10 years (25%). The least experienced researchers, those under two years in, are the most likely to sit out entirely showing up disproportionately both among those who haven’t gotten around to AI yet and those who’ve deliberately chosen not to.

Sector shapes the picture just as much as experience does—healthcare trails every other field, with the lowest overall adoption in the sample. Role matters too, and here the pattern flips: doctoral researchers make up the largest group in the survey overall, but among researchers who use AI regularly—faculty and postdoctoral researchers—take the lead instead, at 50% versus 33% for doctoral researchers. Regular use, in other words, skews more senior than the sample as a whole.

Where researchers reach for AI most is also where they trust it most. For combined regular and occasional use, literature review sees the highest AI use (71%) and transcription the second-highest (58%). Those same two tasks are exactly where trust in AI-generated outputs runs highest too: 34% report high or very high trust in AI for summarizing coded material, and 33% for transcription or data preparation. Interpretation and sense-making, the task researchers hand to AI least (just 8% regular use), is also the task they trust it with least (20% high or very high trust).

That’s not a coincidence—it’s the same instinct showing up twice, once in behavior and once in stated trust. Transcription and data preparation stands out as the clearest case: it’s the one task in the survey where trust in AI’s output outweighs distrust, making it the closest thing to a settled, comfortable use case researchers described.

Given where trust breaks down, researchers were direct about what would change that. Several themes stood out, each pointing to a different layer of the gap between AI’s current capabilities and what research demands of it.

The most consistent ask was for verification built into the tool itself—researchers want AI-generated findings backed by visible evidence and traceable to source material, not asserted on faith. Close behind was a call for human oversight as a structural rule rather than an individual habit: AI treated as a first draft or second opinion, with a human required to confirm before anything counts as a finding. Clearer institutional guidance and stronger data privacy guarantees rounded out the list, often specifically framed as an approved, vetted list of AI tools rather than an open-ended policy.

Together, these asks describe something more specific than “better AI”—they describe rigor as a built-in feature, not an added responsibility for the researcher to carry alone. Researchers have, in effect, already written the requirements. What has been missing is a tool built to them.

Separately, when asked what would need to change in the AI tools, policy, or their field more broadly, a smaller group raised something more fundamental than a feature request—moral and environmental objections to AI as a category, separate from how well any given tool performs. One respondent, a doctoral researcher, put it plainly: they didn’t see their objections changing “until major environmental and economic issues are addressed.”

Government agencies, healthcare organizations, and commercial research settings were examined together as a non-academic segment, separate from the broader academic-dominated sample. Usage here leans occasional rather than routine: 35% use AI for specific tasks now and then, and another 24% have tried it without settling into consistent use.

The same three groups also surfaced in the broader qualitative coding, where their described experiences and concerns diverged by setting:

Data privacy tells a more layered story here than the headline concern suggests. General concern about privacy and confidentiality runs lower than the overall sample—65% somewhat or very concerned, versus 85%. But satisfaction with how their own tool actually handles data privacy and security is far less settled: only 25% are very satisfied, and the largest group (35%) is neutral.

Not every use of AI carries the same risk. A researcher using AI to draft an interview guide is taking on very little; a researcher relying on AI to surface a theme, with no way to trace it back to the data, is taking on a great deal. What researchers in this report are asking for isn’t rigor performed despite the AI tool—it’s rigor the AI tool helps put on the record.

The findings in this report set a bar, and we at Lumivero intend to be measured against it.

For more than three decades, researchers have trusted our research software to hold the infrastructure of rigor: audit trails, transparent coding, a defensible chain of evidence from raw data to finding. As we bring AI into that same environment, we are holding it to the same standard the researchers in this report demand of it:

We are building AI that answers to the researcher, not the other way around. The costs this report documents—hours spent verifying by hand, rigor performed invisibly in side documents, institutions restricting what they cannot audit—are design problems, and we are treating them as ours to solve. Researchers should never have to choose between speed and rigor; our job is to build software where that choice never comes up.

You bring the judgment. Our job is to keep the record.