Congruent use of AI for qualitative data analysis — a message that went viral

Dear Reader,

I am sorry for the long absence. Some of you were already asking whether I am all right. Yes, I am.

In order to manage to write a regular newsletter - there is always more to do than fits in a day - I had put it into my calendar and tried to make a habit of it. Then another regular meeting was put into its place.... and kicked the habit. I am back now. "Newsletter" is a regular recurring event in my calendar again.

What I want to share with you today, is a post that went viral last week. I had pre-planned it to be posted on Saturday at noon because we were invited to a wedding. As LinkedIn's algorithm does not favour if you add an external link inside a post, I thought I check after it was posted an add it to the first comment. By the time I checked, Saturday at 6PM, the post had over 90.000 impressions, over 2000 likes, it was shared over 100 times and had 108 comments 😱. Wow! Since then, LinkedIn showed it to over 450.000 accounts and it has been shared over 550 times.

Here is the text that triggered the virality:

If you're using ChatGPT, Claude, or NotebookLM for qualitative analysis and it feels off — this is why.

There are two different ways GenAI-assisted qualitative analysis goes wrong, and most critiques (and most defending positions) collapse them into one.

The first is technological incongruence: asking an LLM to do something it isn't built for — like expecting a chatbot to "remember" a growing codebook across a whole dataset and apply it consistently. It can't. It doesn't retain a persistent analytic state between prompts, so what looks like coding is often just a plausible-sounding output shaped by the last few pages it saw.

The second is qualitative methodological incongruence — and this one is sneakier. Here the AI workflow reproduces the surface steps of qualitative analysis (codes, themes, quotes) while quietly stripping out the interpretive and reflexive commitments that make those steps meaningful in the first place.

A "theme" needs a researcher's positionality and judgment behind it. What an LLM hands you is a topic bucket wearing a theme's clothes.

In our new paper Kien Nguyen-Trung and I name both problems clearly, then proposes a way through: shifting from "AI as coder" to "AI as interlocutor," with four tenets for keeping the researcher firmly in the interpretive lead.

A lot of people picked up on the sentence: What LLM hands you is a topic bucket wearing a theme's clothes. If you are interested in the comments, click here to read the original post.

The paper I wanted to promote with this post is:

Towards a methodologically congruent framework for GenAI use in nonpositivist qualitative research

It is open access, so everyone can read it. I hope you find the article useful for your work. And if you have any comments, please let me know.