Building Grounded Theory with Generative AI Without Giving Up Researcher Judgment

That problem sits at the center of “Generative grounded theory (GGT): Inductive theory building in the age of generative AI,” a Methods Dialogue article by Bernd Schmitt, Shuyi Hao, Michel Tuan Pham, and Reto Hofstetter. The authors explore how generative AI can support the demanding process of analyzing qualitative data and gaining insight from it. The article demonstrates how qualitative researchers can identify patterns, examine emerging explanations, and develop insights while retaining responsibility for key interpretive decisions.

We spoke with Bernd Schmitt, Robert D. Calkins Professor of International Business at Columbia Business School, and Shuyi Hao, Assistant Professor of Marketing at ICN Business School, to learn more about how generative AI can be incorporated into qualitative analysis and what new possibilities it may open for qualitative research.

Grounded theory and qualitative analysis in general build concepts through sustained comparison with data rather than beginning with a fixed hypothesis. While this makes it particularly valuable for exploring complex phenomena, it also demands considerable analytical effort. Researchers must move between close reading and abstraction, compare fitting and non-fitting cases, revisit earlier decisions, and document how categories evolve. Schmitt and colleagues argue that a central challenge lies in researchers’ limited capacity to process increasingly diverse and extensive qualitative evidence while considering competing interpretations and maintaining traceability.

GGT matters because it treats generative AI not as a shortcut to a finished “theme” or theory but a structured way to use generative AI throughout the analysis process, from identifying patterns in the data to exploring alternative conceptualizations and refining theoretical explanations. The framework is especially relevant to debates about researcher agency, transparency, and whether AI-supported analysis can remain genuinely grounded in evidence. Schmitt reflected on the broader significance of the work: “New phenomena are emerging all the time, and we have more data than ever to study them. The challenge is making sense of this growing volume of information and developing theories that explain what we observe. GGT offers a systematic way to analyze large and diverse datasets, explore alternative interpretations, and develop new theoretical insights. It is not about replacing researchers, but about helping them tackle questions that would otherwise be difficult or too time-consuming to investigate. And doing so systematically and transparently.”

The paper also addresses important concerns surrounding GenAI use: hallucination, omitted evidence, bias, privacy, consent, authorship, and the risk of over-reliance on plausible AI outputs. Hao stressed, “our framework emphasizes clearly defined analytical tasks, explicit links to empirical evidence, systematic consideration of alternative interpretations, and continued researcher oversight throughout the whole process”.

The paper is primarily a methodological proposal, which the authors illustrate through replications and re-analysis.

• GGT turns AI use into a traceable seven-step analytic process. The framework moves from corpus formation and data structuring to coding, clustering, abstraction, theoretical integration, and saturation assessment. Across those steps, the authors specify three guardrails: task-boundedness, evidence anchoring, and comparative orientation. Prompts should correspond to recognizable grounded-theory operations and outputs should remain linked to evidence rather than inviting a model to “create a theory” in one opaque leap.
• The demonstrations show where researcher judgment changes the analysis. In the body-adornment illustration, the initial corpus combined 15 digital-twin interviews with 180 forum discussions, followed by nine additional digital-twin interviews for theoretical sampling. GenAI generated candidate codes, but the researcher evaluated alternative conceptual directions and selected a process model in which self-body misalignment leads to bodily intervention, embodied affect, and restored alignment. Fewer than 12% of early codes were removed as clearly non-descriptive or irrelevant, showing active curation rather than automatic acceptance.
• Structured GGT prompting reproduced established theories more closely than unstructured prompting in two reanalyses. Each condition was run ten times using GPT-5.2. For the published theory of fun, cosine similarity with the original theory was 62.5% under an unstructured benchmark, compared with 81.9% using GGT on original narratives and 83.1% using GGT with synthetic personae. For the theory of special consumption experiences, similarity rose from 59% in the benchmark to 85.4% with original data and 78.4% with synthetic personae. These figures measure textual similarity to the published theories, not the “accuracy” of qualitative interpretation.

Taken together, the illustrations show that the analytic request, theoretical lens, evidence supplied, and researcher interventions can materially change the output. Schmitt reflected on the result: “What we found particularly interesting is that the same data can lead to different theory depending on different theoretical explanations. In our study of synthetic companionship, we found that changing the theoretical lens also changed how the phenomenon was interpreted. This highlights how the theories we bring into an analysis can influence what we see in the data and which aspects we emphasize. It also reminds us of the importance of considering alternative perspectives before settling on a theoretical explanation.”

GGT builds on established grounded-theory procedures and integrates GenAI into a flexible seven-step framework. The seven steps correspond to initial coding, constant comparison, category development, theoretical sampling, integration, and saturation. Prompting is treated as part of the method because wording shapes what the model attends to and what comparisons or abstractions it returns. Researchers can therefore use GenAI in selected steps rather than delegating the full analysis.

The article demonstrates the method in three ways. A body-adornment project shows the full seven-step workflow using forum material supplemented with synthetic interviews. A second project analyzes the same synthetic-companionship dataset under five theoretical lenses. Finally, two published grounded-theory studies are reanalyzed under GGT and benchmark conditions so the generated theory structures can be compared with the originals.

The authors also acknowledge certain limitations. Hao said that “some demonstrations rely partly or entirely on digital twins – not because we believe interviews with real people are no longer needed but simply to demonstrate the method.” The replication test uses similarity to existing theories, which is useful for assessing how closely GGT can reproduce established theoretical structures but does not directly establish the quality of newly developed theories. “Again, we wanted to illustrate the method, as a proof of concept,” said Hao. The authors then highlight that prompts and training data can influence analytical outputs; therefore, maintaining researcher oversight is important as AI-assisted workflows become more autonomous.

For qualitative researchers, one of the paper’s key implications is how researchers and GenAI can work together. GenAI can act as assistant, comparator, co-theorist, or auditor for bounded tasks, but the researcher still decides the question, analytic direction, category relevance, and whether an explanation is theoretically persuasive. That distinction avoids collapsing the debate into either full automation or categorical rejection.

The framework could also improve transparency in qualitative research. If prompts, coding links, rejected alternatives, sampling decisions, and saturation checks are preserved, reviewers and readers can inspect more of the path from evidence to theory. But traceability alone is not enough; researchers still need to critically evaluate AI-generated interpretations and the evidence supporting them.

The next question is therefore less about whether GenAI will enter qualitative research and more about what standards should govern its use. Schmitt reflected on the practical boundary researchers should keep in view: “The biggest risk may not be that AI produces a wrong answer, but that it produces a convincing one too quickly. GGT encourages researchers to question initial explanations, explore alternatives, and ask what evidence might change their interpretation” The paper leaves room for more autonomous systems in the future, but its present framework is built around an accountable human researcher who can challenge, revise, or reject what the model proposes.