Can You Use Gen AI for Qualitative Data Analysis? Here Is a Procedure, Not a Position

Short answer: Yes – if you’re doing certain things under certain circumstances that you have to verify yourself. No, if you do those things by producing this kind of content in an afternoon.

Large Language Models (LLMs) can help us carry out some types of qualitative analysis. For example, descriptive coding using a fixed set of codes would be one task where I would consider involving LLMs – though they still fabricate quotations, get confused about how to run multiple times and move towards very generic analyses. But there’s lots of other qualitative work you’d want to avoid handing to LLMs.

This doesn’t mean we shouldn’t use AI. There are ways we can engage with AI responsibly while avoiding these traps – my advice is outlined here [link to article], and I offer a free micro-course on exactly what that looks like here at BlendedIQ.ai.

So why all this back and forth? In 2025 over 400 qualitative researchers from 30+ countries signed an Open Letter calling for rejection of Generative AI for Reflexive Qualitative Research (the whole thing). Their reasons were based on the assumption that all qualitative research was done exclusively to ‘make sense’ of data which could be seen as being generated through the human mind, rather than engaging in social reality – but as their response points out [link] the argument fails to acknowledge humanity’s history of working collaboratively with tools to think and create meaning.

But what about the people who have been doing this work, including doctoral researchers and working analysts alike — what if we ask them, “Given that these two camps are not talking to each other, could you tell us anything at all about which specific tasks — if any — could potentially be outsourced to a language model when carrying out analysis on your data set for your project?” No one had an answer.

Instead, I asked myself, ‘How would I know whether a big language model might help me carry out some particular tasks?’ I concluded that only evidence obtained locally, produced from my materials under my own configuration, against my own unassisted analysis, would allow me to form a view.

What Can A Large Language Model Actually Do In Qualitative Analysis?

It’s important to consider analysis as a bundle, rather than as individual tasks. Different tasks differ vastly in terms of their checkability:

Descriptive coding – this is the kind of coding I can defend most easily with my own codebook. The model will classify each segment based on how well it fits into the definitions I wrote. It’ll give me an option to specify “no code”, and then when reporting results, for each code, it’ll tell me how many times different coders assigned that code compared to others. That’s good because if we take the average, the thing about the interpretive code where the model failed completely is hidden behind the averaging out.

Summaries and candidate codes – This could be useful as raw data, provided that each piece of data was clearly attributable to its originator.

Candidate themes – These are only acceptable in the sense they’re just candidates. We should ensure all our quotations were found by searching through the transcripts. All these themes need to be binned between grounded (i.e. there actually seems to be some support from the data), plausible but generic (there doesn’t seem to be evidence either way) and unsupported (we don’t have enough evidence). The latter two categories really matter since anything that sounds like it’d fit into any interview for any study is a theme that’s universal across participants and therefore has nothing in particular to say about anyone.

• Interpretation – no delegation can be justified for this activity. You are only responsible for those claims that you yourself originate (model-originated) and trace back to their source, adopting them on your own accorded in your own recorded decisions.

Should I feel safe putting my interview data into ChatGPT, Claude, or Copilot? No by default. There’s far too much hand-waving with these types of instructions. You need to ask yourself these 3 questions, and make sure you can answer yes to all of them:

1. Do I have consent from my participants? Does the way we understand their agreement (as they would) encompass using their data in such a manner?
2. Does our institution permit use of this particular tool for academic purposes – for instance, are there any guidelines published about what kind of data classifications fall under the “approved” list of tools?
3. Does this provider’s Terms & Conditions state clearly if/whether it uses participant inputs for training purposes and whether they retain them? Depending on where you’re located, that may vary greatly based upon plan and tenancy. It really all comes down to checking out your contract!

Anonymization – Anonymization is just a safety net. If we use a consumer chatbot whose transcripts are identifiably human (or worse, have other personal details), we fail the gate.

Enterprise Tenancy & Pseudonymization – Imagine if instead we had an enterprise tenancy where only pseudonymized excerpts from the conversation were accessible to our system. This could work as well, although I think there are many subtleties about how this would be possible that I haven’t considered here yet. Alternatively, what if we trained a completely open weights on our own hardware? There’s no way our model would ever transmit data over the network!

The Local Model – Solving the first problem doesn’t solve the second. Even a local model might produce fabricated results. How can one determine whether the output of AI during qualitative analyses is valid?

Let’s explore three things we should be doing. Each of them will require some practice while taking this course. Let’s begin!

1. The Fabrication Hunt: Ask them for verbatims. Then go hunting through their transcripts looking for things they said that contradict reality (you won’t find any at first — once you’ve caught one, though, you’ll be too scared ever again to paste an unverified model quote into a paper).
2. The Consistency Check: Take the same transcript and prompt and do 3 new chat sessions with different themes. Are your themes dependent on which session you kept? What’s their epistemic status?
3. Traceability Audit: For each proposed theme from the model, double check all of its supporting quotes against their source and categorize each statement as Grounded, Plausible (but Generic), or Unsupported.

In order to pass current academic integrity standards I’d say your statement needs to cover these five points: 1) disclose, 2) own it, 3) verify, 4) protect the data, 5) keep records.

A defensible disclosure statement will include things such as naming the specific tool/verison that was used, stating the tasks that were delegated (and those that weren’t), describing the level of oversight provided and including specifics around the verification process. It could also include details around when and where the model’s reading diverged from yours and what steps were taken at that point – but this might be more involved than just putting together a statement.

It’ll probably help to maintain a decision log of sorts alongside your statement – perhaps a single line entry for every time you contacted the model (see below).

We’ve included an opportunity to make that declaration after completing the whole course, too—so if you take the entire course and then choose this path, I hope you’ll go through our entire rigorously structured set before you say “no.”

In short: This course doesn’t hold any positions for/against (with confidence and bibliography rather than with evidence), but we recognize the need for people to declare their own affirmation of non-use when there’s otherwise little to talk about in our current climate. So please do complete the whole course first!

So what exactly does the free course cover? Using Large Language Models for Qualitative Data Analysis is now available as a free, self-paced ‘micro’ course hosted at BlendedIQ.ai. All learning takes place using our purpose-built synthetic interview corpus (we never see your data!) consisting of six modules plus a bonus philosophy seminar!

The tooling & data safety section covers all that and more! Here’s my plan:

1. Tooling & data safety – What the model does, and the three part data gate to catch your own fabrication
2. Unassisted baseline – Code a transcript by hand then seal your reading in a dated memo (no model contact yet)
3. Deductive work flows – Your model is now your code book with measurable results per code leading you to your evidence based delegation verdict
4. Inductive Workflows – Themes on Trial (with full Traceability Audit)
5. Abductive Workflows – Turn the Model’s Conventionality Into A Mirror That Surfaces Your Assumptions
6. Academic Integrity – Your Disclosure Statement & Conditions Note, Drafted From Your Own Evidence

Bonus Module 7 on the Philosophy Behind the Debate – Five Positions Mapped, And Your Own Position Statement.

Who Is This Course For? Doctoral researchers and their supervisors in the Social Sciences and Management; Qualitative researchers looking at considering using these techniques for the first time; Methods Instructors wishing to get some teachable, evidenced based workflows into their curricula; AND Skeptics (we’re really excited to have you here!)

You’ll Need A Browser & Spreadsheet Application & Access To Any Major AI Chat Tool.

Frequently asked questions

Can we use ChatGPT (or other LLMs) to help analyse our qualitative interviews?

You can only use LLMs if you have access to some sort of ‘data gate’ which ensures all outputs are checked against the raw sources. Otherwise unverified model output isn’t really ‘analysis’.

I’m seeing models fabricating quotes from the transcript – how does this work?

The models don’t search your transcripts and then extract and format quotes as they appear, instead the models will generate possible answers based upon everything in their training set. This means that the chance of the model generating fake quotes that sound like authentic quotations from a transcript increases as the length of the transcript increases.

Do I need participant consent to apply AI techniques to the data?

Your AI processing must be covered by an agreement that the research participants would reasonably consider appropriate. For example, most informed consent forms give permission to do basic statistical analyses using anonymised datasets. You’ll need to ensure that your choice of AI processing falls inside those bounds. If there’s room for ambiguity, ask your ethics committee in writing whether they approve of what you’re doing.

Does anonymizing my data satisfy this condition?

Not quite. We’re talking about what you can safely do (a ‘permission’) given a particular set of safeguards. You don’t get to decide whether anonymisation will work for you — we’ve written on how often it doesn’t. But if that’s all you did, well done!

How should I include my disclosure of using AI tools in my thesis?

Give the name of your tool(s), their versions and tasks performed on them. Also indicate who oversaw the process and ways of verifying outcomes are discussed. This should be supported by relevant figures. A helpful practice is to maintain a ‘decision log’ as you proceed, which makes writing up easier later.

Is this course actually free?

Yes! Simply go ahead and enrol here at blendediq.ai. Then dive into our course overview. Bring along your questions and scepticism — that’s exactly what the course was designed for. If you are a lecturer or professor, the teaching resources are free as well. If you like them, buy me a beer at the next conference.

Key takeaways

The AI Debate In Qualitative Research Offers Positions. Defensible Practice Requires Procedure. Delegation Is Earned Per Task, Evidenced Locally On Your Own Materials And Revocable When Conditions Change.

Your Unassisted Reading, Dated Prior To Any Model Contact, Is The Instrument Every Later Check Depends On.

• Verify everything –quotes searched, agreement per code, themes traceable back to source. • Refusal with evidence is just as valid as use with evidence • Enrol for Free here http://BlendedIQ.ai Using LLMS for QDA.