Key takeaways

• The discussion on using AI in qualitative research often pits two unhelpful approaches against each other: a total ban vs. unfettered AI automation.
• AI tools add value in small, targeted ways. They can improve workflows, support better research design, and help explore and validate evidence.
• Ultimately, impact depends on how improvements in efficiency are used. AI can deepen research insights or just increase output, shaped by incentives and client demands.

The debate on the use of AI in qualitative research is often framed by two extreme positions.

On the one side are the methodological purists. This group includes the scholars who, in a recent open letter to the Qualitative Inquiry journal, argued that generative AI has no place in certain forms of qualitative research, as it is unable to generate meaningful insight—its outputs are statistical predictions, untouched by context or experience.

On the other side are those advocating for automation at points throughout the research process—from data collection and analysis to reporting. This side is more diverse in its makeup and background, comprising professionals who commission and/or develop AI solutions for qualitative research yet may have little understanding of what such research is and entails. It also includes some applied researchers with social science backgrounds using AI due to lack of resources and/or narrow client expectations/requirements.

This all-or-nothing framing is not helpful. A total ban on the use of AI in qualitative research is unrealistic. Given the rapid spread of AI tools, soon that may be like asking someone not to use the Internet. But neither should researchers be using AI tools indiscriminately. These are important issues for food system and global development research, which stand to benefit from some forms of AI automation and also face tenuous funding.

In this post, I focus on the spaces and possibilities between the two extremes of the debate, arguing that AI in various forms can be integrated more thoughtfully and more responsibly at key moments during the research process, perhaps in less obvious ways than what is suggested by each extreme side in this debate.

It is also important to acknowledge that discussions on the role of AI in qualitative research are not only about analytical rigor and that decisions on if and how to use AI are not something that researchers alone can make.

Echoes of an earlier research debate

It is useful to note that a version of this argument happened 30 years ago, when digital applications—so-called computer-assisted qualitative data analysis software (CAQDAS)—including NUD*IST, ATLAS.ti, and later NVivo and MAXQDA, entered qualitative research. Some researchers argued that the new tools would push qualitative work toward a uniform mode of analysis—coding everything, grounded-theory-style, while marginalizing other equally valid and useful qualitative approaches. Others pushed back, arguing such criticisms were largely unsubstantiated and kept alive by careless citation.

In retrospect, both sides were partly right. Many research traditions never adopted CAQDAS. But scholars who did, including many of the signatories of the Qualitative Inquiry letter, contributed to the development of standards that enabled the meaningful integration of CAQDAS in research. In the process, however, coding did become the prevalent mode of qualitative analysis in applied research.

The CAQDAS parallel reminds us that there are different traditions in qualitative research and that such debates are not simply about analytical rigor. They are also about professional identity and about how certain approaches come to be regarded as more “scientific” than others.

There is more at stake with AI. CAQDAS streamlined certain aspects of analysis. AI automation threatens to hijack the research process itself.

But wait, what do we mean by AI?

As already noted, many AI enthusiasts developing tools for data collection or analysis have a limited understanding of qualitative research, the rich traditions that it encompasses, and the processes that it entails. By the same token, many methodological purists may not fully understand “AI.”

My hunch is that what many methodological purists understand as AI are popular browser-based large language model (LLM) interfaces such as ChatGPT, Gemini, or Claude.

This matters, because how we access AI shapes what it can do. The browser chatbot is one option and, many would argue, the most limited computationally (and thus methodologically) and in terms of information security. More useful alternatives include accessing LLMs via application programming interfaces (APIs); retrieval-augmented generation (RAG) approaches, as exemplified by Google’s NotebookLM; using AI on the desktop (e.g., the Claude app or Microsoft CoPilot 365); and CAQDAS software with AI features.

It is also worth pointing out that the term “AI” refers not to a single technology but to many. The LLMs behind these interfaces are only the most visible. Speech-to-text systems generate transcripts. Machine translation handles multilingual data. Older natural language processing methods such as topic modeling and sentiment analysis predate the current wave and still do useful work. Lumping these diverse tools under a single term makes meaningful conversations challenging.

Benefits of using AI for qualitative research

In my experience, there are three benefits AI can bring into the qual research process:

Small cumulative gains across the workflow. Checking that interview questions actually address research questions. Generating clean transcripts in multiple languages. Drafting codebooks for human refinement. Maintaining better project documentation. These are not headline-grabbing applications. They are the kind of incremental support that frees researcher attention for the most demanding aspects of qualitative work.

Better research designs. Used carefully, AI can help us analyze secondary evidence faster, allowing us to refine our approach before data collection and then to ask the right people the questions that cannot be answered from existing data sources, instead of using interviews solely for fact-finding. This approach enables more methodological options; in particular, it gets us closer to actually implementing mixed-methods designs, which work well on paper but are very challenging to realize. This also means that we can use individual interviews or focus group discussions more strategically, getting more out of fewer interviews (and making better use of our participants’ time). Except for cleaning interview transcripts, standard browser-based AI chatbots can perform many of these tasks, especially when provided with information on the study context and approach.

Meanwhile, the trend towards AI-conducted interviews—which enable scaling up the interview process, but with far less contact between researchers and subjects—carries a number of risks. It might collect less useful data—or data better gathered via other means—even across a large number of interviews. It may also discourage participation among subjects once the novelty wears off.

New ways to explore, validate, and share evidence. Retrieval-augmented systems can allow researchers to query their evidence base and triangulate some findings (e.g., “How many participants said something close to this?”), consider findings from different perspectives, and identify blind spots in their own understanding and analysis. AI-generated translated summaries of research materials can support coordination when data collection/analysis is conducted in multiple languages.

Using such AI tools effectively requires investment in terms of researcher time and institutional support, albeit to differing degrees. And all of it is meaningless if the gains flow in the wrong direction—towards doing more of the same (going for more bids, delivering more projects)—instead of creating more space to think and reflect, deepen analyses and learning. That is, unfortunately, a real possibility in the current climate.

The bigger questions

Whether AI improves qualitative research or degrades it depends as much on the technology and how it is leveraged as on how the time and resources it saves are used. If an AI tool saves six hours, that time can be used to improve the quality of the work or to increase the quantity to make offerings more competitive. These time savings can quickly become something that clients expect by default, perversely ratcheting up time pressure on users.

These are not questions researchers can answer on their own. For applied researchers, such as monitoring evaluation and learning (MEL) professionals, part of the answer lies with commissioners’/clients’/funders’ expectations and what their own organizations reward (epistemic standards also matter, but these take more time to crystallize).

If the demand is for more data points, faster turnaround, and lower costs from qualitative research, then AI will become a set of tools for producing more output and less insight (and one hopes that we will realize and course-correct quickly). If the demand is for more contextual, more validated, more thoughtful work, then AI can become a useful partner.

Evangelia Berdou is AI & Digital Integration Lead with Agulhas Applied Knowledge. Opinions are the author’s.