Researchers at the University of Missouri built a tool to test whether AI can do one of the trickier jobs in qualitative research: reading interviews and deciding what themes run through them. Their audit found the AI produced at least one matching topic for 11 of the 12 themes human researchers had identified, yet often missed what those themes actually meant

2

1

.

Source: Digital Trends

Source: Digital Trends

The tool is a six-stage framework, the Alignment and Usefulness Rating Audit, and the study appeared in the International Journal of Qualitative Methods

2

1

. Qualitative researchers, the kind who work through interviews, focus groups and open-ended survey responses looking for themes, have long done that reading by hand, and it eats time

1

. The audit is meant to help them judge when an AI's topic output is actually relevant and useful

2

1

.

What the test found

The team ran the audit on 32 interviews with nursing educators, clinicians and students about how they use generative AI

2

1

. They compared topics produced by BERTopic, a text-analysis system, against 12 themes human researchers had previously identified. On 11 of the 12, the system produced at least one sufficiently similar topic

2

1

.

A match, the researchers found, did not mean the model understood a theme. It sometimes merged several distinct themes into one broad topic, and sometimes split a single theme into several smaller ones, blurring the point the interviewees were making

1

. The team also found only limited evidence that the model added new insights of its own, beyond what the human analysts had already surfaced

2

1

. The test used BERTopic, a topic-modeling system, rather than a conversational chatbot such as ChatGPT or Gemini

2

.

Who ran it, and what's next

The team itself spans journalism, data science and nursing, and it says that mix is what lets it judge where human judgment remains essential

1

. Named members include Chau Tong and John Robert Bautista, both assistant professors, and doctoral researcher Temitope Eyinla

1

.

The authors themselves flag the limits: 32 interviews is a small, single-domain test. More work is needed, they say, to know how well the audit and the findings hold for other AI systems, subjects and settings

2

1

. They think the approach could eventually support research in healthcare, education and social science, wherever interpreting context matters

2

1

.

Their own read is that the two can work as complements. "AI can help researchers see patterns across large amounts of information, but the real question is what those patterns tell us about people's experiences," Tong said

1

. Per Bautista, interview conversations rarely stay on one topic start to finish, which is where sorting help comes in

1

. "But human expertise remains central to interpreting the results," Eyinla said

1

.

© 2026 TheOutpost.AI All rights reserved