A 60-year-old Grounded Theory dispute resurfaces as an AI problem

A 60-year-old Grounded Theory dispute has just resurfaced as an AI problem

I have been reading Schmitt et al.'s new paper: Generative grounded theory ( GGT ): Inductive theory building in the age of generative AI

doi.org/10.1002/jcpy.70036 . Journal of Consumer Psychology.

Their proposed workflow is sophisticated: GenAI generates descriptive codes, clusters them into categories, develops higher-order concepts and eventually contributes to theory construction. The researcher is supposed to remain responsible throughout.

This made me think again about the old dispute between discovery and construction in Grounded Theory. The title of Glaser and Strauss's 1967 book was deliberate: The Discovery of Grounded Theory. Later, Kathy Charmaz challenged the assumption behind this language. She said: Categories and theories are not simply waiting inside the data to be found. They are constructed through the interaction between researchers, empirical materials, questions, theoretical resources and context.

GenAI makes this distinction between finding something in the data that presumable already existing, and constructing meaning through the process of engaging with data very relevant.

If an AI tool can identify something that objectively sits in the data just wating to be discovered, then it becomes an obvious efficiency tool. Then it makes sense to ask it to: • Find the codes • Cluster them • Identify the relationships • Build the theory

If you take Charmaz's perspective, then every one of those outputs is already a construction. A model cannot reveal the categories in the material. It generates one possible account, shaped by the prompt, model, system instructions, data preparation and everything that came before. Human oversight at the end does not turn this into something objective, the only one possible solution that exists.

I find it highly problematic to let an LLM do the coding and the researcher's role consists of evaluating the outcome. This is not a question of epistemology but of ontology. You need to believe that there is something objective to be found in the data that can be discovered by a technical entity. If - as researcher - you ask AI code the data, you are not part of the process and simply have no idea why something was coded in what way.

We are not talking about coding 1 or 2 interviews where you can still check everything and be part of it. Basically you are confronted with an output in the end. If you tried it, you know that it is very difficult to take ownership. If someone asked you a question about the data, you are lost for an answer.

When researchers outsource coding to an AI and limit their role to reviewing the output, they are not just automating a tedious task; they are outsourcing the epistemic work.

We have seen this now in many articles that an LLM does not possess embodied experience, contextual immersion, or historical situatedness. Outputs are generated by parsing tokens and predicting linguistic patterns based on learned associations, not through lived understanding.

When an LLM processes data to create an "evidence-linked coding table" and subsequently clusters these into provisional categories, the researcher is robbed of the messy, vital struggle of engaging deeply with the material.

Reviewing an LLM's categorical output is fundamentally different from building those categories organically. Defending the nuanced origins of a specific code becomes nearly impossible when the researcher is not part of its genesis.

Therefore, my slogan still stands: Stop coding. Start talking.

I strongly advocate that researchers need to remain involved in the process of data analysis rather than outsourcing it - even if coding data has become technically much more feasible than it was just 6 months ago.

The question should not be: How much of coding can we automate? But: How can AI support the researcher's engagement with the data without taking over?

This brings me to the methodological congruence article

doi.org/10.1080/13645579.2026.2703248 that I recently published with Kien Nuygen-Trung:

Methodological congruence cannot simply be established by preserving the names of analytical procedures like open, axial and theoretical coding while a) not reflecting adequately on the ontological position and b) by changing who—or what—performs the epistemic work within the described analytical procedures.

Something to reflect on this week. Until next Monday,