Delve

Is AI Analyzing, or Just Chatting? A Review of Nguyen and Welch

Paper review: is AI analyzing, or just chatting?

Paper review: in this series we summarize academic research on qualitative analysis.

A recent paper by Duc Cuong Nguyen and Catherine Welch argues that generative AI is unsuited to qualitative data analysis. It isn’t hard to see how they got there. The first forays into AI by qualitative analysis tools were easy to critique. Promises of “qualitative insights in minutes instead of weeks,” or of analysis done end to end, abandoned the philosophy behind the methodologies they claimed to automate. The authors write from an interpretivist perspective, where there is no objective standpoint and no single correct answer. What exactly is AI doing, then, when it produces themes with one button click?

The paper: Generative Artificial Intelligence in Qualitative Data Analysis: Analyzing, Or Just Chatting? by Duc Cuong Nguyen and Catherine Welch (Organizational Research Methods, 2026).

They propose four criteria for judging AI as a research tool: factual accuracy, reliability, transparency, and ethics. I agree with a lot of what they find. A research community should validate new tools before adopting them, and its role is to be critical when new technologies are pushed into research. And the stakes are high. Research tools, as the authors put it, “are not passive intermediaries; they are constitutive of the knowledge they help produce.” In qualitative research that cuts deeper, because the data “are imbued with social meaning that is difficult to access and can easily be misunderstood.”1 A tool that misreads that meaning shapes what a field comes to know.

The authors note that, “unlike many proponents of the tools (notably Qual-AI providers), we have no financial interest in encouraging its adoption.” I can’t say the same. I’m the founder of a qualitative analysis tool that builds AI features, so I stand to benefit if researchers adopt them. Skepticism of the people promoting these tools is warranted, and in the spirit of reflexivity, that includes me.

The criteria are the right place to start, and they capture much of what went wrong with those early tools. Looking ahead, I disagree with some of the criteria, and on others I think the risk is far more insidious, in ways the paper raises but doesn’t dwell on. Let’s go through them one by one.

Factual accuracy

This is the easiest criterion to stand by, because everyone knows about hallucinations. The authors’ point goes deeper: LLMs are built to produce text that looks right. As one source they cite puts it, the output “does not have to be correct, it must appear correct.”2

But I think hallucinations are becoming the least relevant part of this. They happen less and less, and software built around the models can reduce them further. The authors acknowledge that approaches like retrieval-augmented generation (RAG) can raise the odds of accurate output, though they argue it can’t guarantee it.3

What worries me more is the inaccuracy we can’t see. The authors make a good case that AI-generated themes are currently poor: topic summaries, often misleading, often inaccurate. We’ve seen this ourselves. When we tried building AI tools that create themes for you, we didn’t think the results were good enough to launch to the research community.

Then there’s the query-based analysis camp, who argue you can talk your way to themes by iterating with AI. The authors tried this: after eight rounds of prompting, ChatGPT began “producing the appearance of the kind of results that we were explicitly soliciting.” They object that the process never converges and eats up time, and I partly agree: sometimes you spend more time reading synthetic analysis than the data itself.

But notice what happened in their test: the output got better at looking like what they wanted. That points to something scarier than factual errors. You can catch factual errors. What you can’t catch is epistemic capture: the slow handover of your judgment to the model. You still decide which readings to accept, but the model is supplying the readings, and it has learned which ones you accept. Over time, its sense of what your data means becomes yours.

When you iterate with AI, is it getting better because it has found an interesting concept? Or because it has learned to phrase things in a way you’ll accept? The authors draw the comparison themselves: models tailor outputs to user preferences, “much like how social media algorithms curate content recommendations.” Hallucinations were the first and most visible version of this problem, and the most fixable. The harder version comes when the errors are no longer detectable, because the model has become very good at learning what we want to hear.

There are already stories in the news of people going down deep rabbit holes with AI.4 But this is a gradient. Where along it do we notice we’re in the rabbit hole? And when AI helps us create knowledge, is it what we would have created without it, or are we being captured by its thinking?

The risk is sharpest for query-based analysis, where the analysis is mostly a conversation with AI about your data, with no coding behind it. There, the authors found the AI also skips parts of the data. Better models and software may well fix that. Epistemic capture is different: as models get better, they get more convincing, so the risk grows rather than shrinks.

So I’d propose something beyond factual accuracy, or on top of it as a baseline. The questions I’d ask instead: Is the analysis capturing the human experience? Is it capturing the researcher’s subjectivity? Or is it converging on what the AI would say?

That last question matters most to me, because the stakes are collective. One researcher captured by AI produces one weaker paper. But if a whole field analyzes its data through the same few models, nudged toward the same framings, every paper could pass review while the field as a whole narrows, and no single study would reveal it. The authors gesture at this when they warn that widespread adoption could threaten “the diversity of research approaches, leading to a homogenization in knowledge production.” I’d put it more bluntly: are we losing something collectively when we all do this?

Reliability

I find their reliability argument one of the weaker parts of the paper.

They argue AI can’t be trusted because it’s probabilistic, so the same prompt produces different outputs. But that assumes scientific tools need to be deterministic, and many aren’t. Humans aren’t either: we produce different readings in different situations and at different times. The authors quote Karjus making this point, that AI’s limitations “are not categorically unique to machines and also apply to human analysts.”5 They call it a false equivalence, but I find it quite relevant.

There’s also a tension in their own framing. They note that interpretivists “reject traditional positivist criteria of validity and reliability,” then make “reliability (i.e., stability, dependability)” one of their criteria for judging AI. Their likely answer is that reliability applies to the tool, not the analyst. Maybe that holds up, or maybe I’ve misunderstood them. But using the same word in two quite different ways within one paper weakens the argument.

Dependability may be a better criterion. It’s the interpretivist counterpart to reliability, from Lincoln and Guba,6 and it asks whether a research process is consistent and traceable, rather than whether it produces identical results. The authors even use the word in their own definition of reliability. By that standard, the question isn’t whether AI output varies, but whether you can trace how it was produced and why it differs. That leads straight to transparency.

Transparency

Transparency matters especially in qualitative research. In quantitative work, others can rerun your numbers. In qualitative research, nobody can redo your fieldwork or your reading of the data, so your methods are how your findings travel: readers judge your conclusions by your account of how you reached them. Without that transparency, the academic community has a hard time communicating ideas to one another at all. And “I used AI” can mean anything from searching for a quote to generating every theme.

The tools that most obviously fail it are the one-button solutions that promise to do qualitative analysis end to end. ATLAS.ti markets “qualitative insights in minutes instead of weeks,” and AILYZE promises results “30x faster.” Tools like these fail the transparency test almost immediately.

With LLMs, transparency has two layers: transparency into the model itself, and transparency into the tool built around it.

Start with the model. I fear many people using AI for qualitative analysis don’t understand what’s happening. If you load your whole dataset into ChatGPT and ask for a thematic analysis, it isn’t doing thematic analysis. It’s predicting, word by word, what a thematic analysis report would look like. The authors put it plainly. These tools “are not searching, coding, or undertaking interpretive or theory-based (thematic) analysis with the user’s data.” They are “mimicking forms of human text that are commonly associated with the process of analysis.”

Then there are the qualitative analysis tools that incorporate AI. I don’t know the inner workings of every tool, but most combine ordinary deterministic software with calls to an LLM. I know this because I built one. Delve’s autocoder doesn’t hand your transcript to an LLM and say “please code this.” It works through the transcript in small sections, close to line by line, and has the AI apply your codebook to each one. That forces the AI to grapple with all of the data, not just the parts that are probabilistically easy to handle, which addresses one of the gaps the authors found.

One-click, end-to-end solutions, which claim to do a full thematic analysis without any human input, are a different story. I find them misguided and opportunistic. We tried building something similar, and it doesn’t work, and the authors show clearly why these tools fail. But even if they worked, they’d be non-transparent twice over. They don’t tell you how they create themes, a complaint the authors share when they note that developers “do not disclose their prompting techniques, design choices or RAG architecture.” And the LLM underneath is opaque too.

That first layer doesn’t have to be opaque. These tools could keep a real audit trail of everything they did. The second layer is harder: neural networks remain largely a black box, even to the people who build them.

Still, humans are also, to an extent, a black box. We attempt to be reflexive, but in the end you have to trust a researcher’s memos and notes, and believe they genuinely tried to be self-aware and self-critical. Couldn’t an LLM do the same: write memos and explanations, and refer back to them? The authors reject this comparison, and their counter has weight: the errors are different in kind. Humans, for all their biases, can introspect, reflect, and correct their initial conclusions. I’d add that an LLM’s explanations may not reflect how it actually produced an output. Still, I think the comparison is worth grappling with. Not being fully predictable or transparent may not be disqualifying on its own. With a real audit trail, the right use cases, and a scalpel-like approach to how LLMs are used, can they be transparent enough?

Ethics

Some of the ethical concerns the authors bring up are hard to argue with. The immense energy these models consume is hard to justify at a time of climate change.7

The one ethical concern I’d push back on is data leakage, the risk that what you upload ends up in a model’s training data.8 It should be a concern, but it can be addressed with standard security and legal practices. What goes into training data is, at least for now, a company’s choice, and providers like OpenAI offer enterprise agreements that commit them not to train on customer data. That’s not to say we should trust them completely. These are companies that trained their models by scraping vast swaths of the internet without permission. But they have a strong financial incentive to honor these agreements, since breaking them would mean serious legal consequences and the loss of their enterprise customers.

Conclusion

This is a great paper, and it asks the right questions. It warns that if we don’t think about AI critically, there are forces fully intent on pushing it into qualitative analysis anyway.

However, I think the authors are too quick to condemn AI entirely. They’re right about end-to-end solutions, which fail on accuracy and reliability alike. But can AI be used as a scalpel, in ways that support accuracy, dependability, and transparency rather than undermine them? I think that question deserves more exploration, and I don’t think we can write it off as quickly as this paper does.

Where I’m most concerned is factual accuracy. Not the minutiae of hallucination, but the larger problem of epistemic capture: AI dulling our collective ability to come up with new concepts. That alone may be a reason to reject AI altogether. But I think it’s still up for debate.

Go deeper: Generative Artificial Intelligence in Qualitative Data Analysis, Nguyen and Welch, Organizational Research Methods, 29(1), 3–39, 2026. It is open access.

Related guides from Delve:

Try Delve AI for Qualitative Analysis | Start a Free Trial

Cite this article

Limpaecher, A. (2026, October 9). Is AI Analyzing, or Just Chatting? A Review of Nguyen and Welch. https://delvetool.com/blog/ai-analyzing-or-just-chatting

Footnotes

  1. The authors cite Maxwell (1992), Understanding and Validity in Qualitative Research, in the Harvard Educational Review. ↩

  2. The quote is from Thornton (2023), A Special Delivery by a Fork: Where Does Artificial Intelligence Come From?, in New Directions for Evaluation. For the related point that LLMs are indifferent to the truth of their output, the authors cite Hicks, Humphries and Slater (2024), ChatGPT Is Bullshit, and Narayanan and Kapoor’s book AI Snake Oil (2024). ↩

  3. On RAG, the authors cite the original paper, Lewis et al. (2020), Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, along with two surveys, Ni et al. (2025) and Zhao et al. (2024). ↩

  4. They Asked an A.I. Chatbot Questions. The Answers Sent Them Spiraling. The New York Times, June 13, 2025. ↩

  5. Karjus (2025), Machine-Assisted Quantitizing Designs: Augmenting Humanities and Social Sciences with Artificial Intelligence, in Humanities and Social Sciences Communications. It is open access. ↩

  6. Lincoln and Guba (1985), Naturalistic Inquiry, Sage. Nguyen and Welch cite the same book for the term “audit trail,” which they list among the things all qualitative researchers care about, whatever their paradigm. ↩

  7. The authors cite Luccioni, Jernite and Strubell (2024), Power Hungry Processing: Watts Driving the Cost of AI Deployment? ↩

  8. The authors’ source on data leakage is Lukas et al. (2023), Analyzing Leakage of Personally Identifiable Information in Language Models, from the IEEE Symposium on Security and Privacy. ↩