Can Generative AI Create Coding Categories Automatically?

Can AI Discover Coding Categories Without Being Given a Codebook?

You have collected interview transcripts or open-ended survey responses, but you have not yet developed a coding framework. Instead of defining categories manually, could you ask generative AI to examine the responses and identify the categories that emerge?

A model might return labels such as "limited institutional support," "technology-related frustration," and "perceived learning benefits." The results may appear organized and analytically convincing, particularly when accompanied by supporting quotations.

However, producing category names is not the same as establishing that those categories adequately represent the dataset. The central question is whether AI can contribute to category development without prematurely imposing an interpretation on the research material.

Yes, AI Can Propose Categories, but It Cannot Validate Them Automatically

How Generative AI Can Assist With Developing Coding Categories

What Does It Mean to Generate Coding Categories Automatically?

Automatic category generation involves asking an AI system to examine research material and propose conceptual labels without supplying a complete list of categories beforehand.

For example, a researcher investigating teachers' experiences with educational technology might provide several interview excerpts and request recurring concepts that could be useful for qualitative coding.

The model might suggest:

• Insufficient technical training
• Unreliable internet connectivity
• Increased lesson preparation time
• Improved instructional flexibility
• Limited institutional support

These labels could provide a starting point for analysis. However, the researcher must determine whether they represent distinct concepts, whether important experiences are missing, and whether the labels accurately reflect the participants' statements.

In this context, "automatically" describes how the model produces suggestions. It does not mean that category development has been completed without analytical decisions.

How Is AI-Generated Category Development Different From Applying Existing Codes?

Category generation and category application are related but distinct activities.

The distinction matters because a model can apply a poorly designed framework consistently. Conversely, it can propose useful categories but assign individual passages to them incorrectly.

Once categories have been developed and approved, researchers can separately investigate whether AI can apply a human-developed codebook to the remaining data.

Are AI-Generated Categories the Same as Inductive Codes?

Not automatically. Inductive coding generally involves developing codes through engagement with the data rather than applying a fully predetermined framework. AI-generated labels may resemble inductive codes because they are proposed in response to supplied material.

Nevertheless, the model brings learned linguistic patterns and conceptual associations to the task. Its suggestions may reflect familiar ways of describing a topic rather than distinctions that are especially meaningful within the particular study.

For example, an AI system might label several statements as "resistance to technology." Closer examination could reveal that participants are not resisting technology itself. They may be questioning mandatory implementation, inadequate training, or unrealistic institutional expectations.

Accepting the model's initial category could therefore introduce an interpretation that participants did not express.

Parfenova and colleagues (2025) examined inductive text coding using six open-source language models and human experts. Their findings indicate that coding difficulty matters and that model-generated labels can differ from expert judgments, particularly for complex material. This supports evaluating AI-generated suggestions rather than assuming that plausible labels are analytically appropriate. Their published study provides a relevant empirical comparison.

What Is the Difference Between a Code, a Category, and a Theme?

These terms are sometimes used interchangeably in casual discussions of qualitative analysis, but they can represent different analytical functions.

Consider three hypothetical interview statements from teachers:

• "Nobody showed us how to use the new platform."
• "We were expected to teach online without proper orientation."
• "I learned the system by watching tutorials on my own."

The researcher might initially assign codes such as "absence of training," "lack of orientation," and "self-directed learning." These could be organized under a broader category such as "professional preparation for technology use."

A theme might subsequently develop around how teachers were expected to assume responsibility for adapting to institutional technology changes without adequate organizational preparation.

However, not every qualitative approach requires this hierarchy. In reflexive thematic analysis, Braun and Clarke emphasize that themes are developed through interpretive engagement rather than simply created by combining similar codes. Their methodological comparison explains why different approaches to thematic analysis should not be treated as interchangeable.

Can AI Decide How Many Coding Categories a Study Needs?

AI can propose a number of categories, but there is no universally correct number that can be determined from the dataset alone.

If a researcher requests five categories, a model may combine several distinct concepts to satisfy that instruction. If the researcher requests twenty, the model may split closely related ideas into unnecessarily narrow groups.

Category granularity refers to how broad or specific the categories are. Broad categories simplify organization but may conceal meaningful differences. Narrow categories preserve detail but can create overlap and make the framework difficult to apply.

Researchers should determine the appropriate level of detail by examining the research question, conceptual differences in the data, and intended analytical use.

How Can Researchers Identify Overlapping or Redundant Categories?

AI may generate labels that sound different but refer to substantially similar experiences.

For example, "insufficient training," "limited technical preparation," and "lack of professional development" might overlap. Whether they should be combined depends on their definitions and the distinctions present in the data.

Researchers can ask AI to compare category definitions, identify potential overlap, and present passages that challenge the proposed boundaries. However, the decision to merge or separate categories should be based on the analytical purpose and actual evidence.

A useful test is whether two categories can be distinguished using explicit decision rules. If researchers cannot explain why a particular passage belongs to one category rather than another, the framework may need refinement.

Can AI Miss Categories That Appear Only Rarely?

Yes. A model asked to identify the most common concepts may prioritize frequent or linguistically prominent ideas while overlooking less common experiences.

Imagine that most students discuss internet connectivity and learning platforms, while a small number describe accessibility barriers affecting students with disabilities. A frequency-oriented categorization may underrepresent the accessibility concern.

Its low frequency does not necessarily make it analytically unimportant. Some research questions concern precisely those experiences that are uncommon, marginalized, or inadequately represented by broad categories.

Researchers should therefore examine outliers, minority perspectives, contradictory cases, and statements that do not fit the initial framework. Category completeness should not be judged solely by how much of the dataset can be placed into a few popular labels.

Why Might AI Generate Different Categories From the Same Data?

Generative models can produce different outputs across runs, even when the input appears unchanged. Differences may arise from generation settings, model updates, input ordering, prompt wording, and other aspects of the processing environment.

One run might propose "institutional support" as a broad category. Another might separate "administrative support," "technical assistance," and "professional development."

These differences do not necessarily mean one output is correct and the other is wrong. They demonstrate that category generation involves choices about abstraction and interpretation.

Researchers should examine whether coding changes across repeated prompts and retain records of the versions used during framework development.

How Should Researchers Develop a Codebook From AI-Generated Categories?

A list of category names is not yet a usable codebook. Each approved category should be sufficiently defined for its intended application.

A practical codebook may contain the category name, operational definition, inclusion criteria, exclusion criteria, illustrative passages, and guidance for borderline cases.

For instance, a category called "institutional support" could be too broad unless the researcher clarifies whether it includes financial assistance, technical help, administrative communication, or professional development.

Researchers should also retain a record of why categories were introduced, merged, renamed, or rejected. This documentation helps distinguish the model's initial suggestions from the researcher's final analytical decisions.

When a coding framework changes, previously coded material may need to be reconsidered. Otherwise, early responses could have been classified using definitions that differ from those applied later.

How Can Researchers Evaluate AI-Generated Categories?

Evaluation should focus on conceptual and methodological adequacy rather than the attractiveness of the category names.

Researchers can examine whether each category is grounded in identifiable passages, captures a meaningful distinction, contributes to the research question, and can be differentiated from neighboring categories.

They should also investigate what the proposed framework fails to represent. A framework that classifies nearly every response may still distort the data if it forces distinct experiences into inappropriate categories.

For structured content analysis, researchers may subsequently evaluate how reliably the finalized framework can be applied. For interpretive approaches, the quality of category development may depend more on analytical depth, reflexivity, and coherence with the methodological orientation.

As Braun and Clarke (2021) argue, different approaches to thematic analysis involve different assumptions about what constitutes quality. A coding framework should therefore be evaluated using criteria appropriate to the methodology rather than a universal checklist of numerical agreement.

Developing Coding Categories From Teachers' Experiences With AI

Step 1: Examine the Original Interview Material

The researcher first reads the transcripts to become familiar with the participants' experiences. The researcher then selects appropriately prepared excerpts for an exploratory AI-assisted exercise.

Three illustrative statements are:

• Teacher A: "AI helps me prepare lesson activities, but I still have to check whether the examples are correct."
• Teacher B: "Our school encourages us to use AI, but nobody has explained what we are allowed to upload."
• Teacher C: "I can prepare materials faster, although sometimes the suggested activities do not match my students' abilities."

Step 2: Request Initial Category Suggestions

The researcher asks AI to propose descriptive codes and possible broader categories, identify supporting passages, and avoid making claims not supported by the excerpts.

The following is an illustrative output, not an actual model evaluation:

Step 3: Evaluate the Proposed Categories

The researcher notices that "verification of AI outputs" and "instructional appropriateness" may be related, but they are not necessarily interchangeable.

Teacher A describes checking whether information is correct. Teacher C describes whether instructional activities suit students' abilities. Combining both under "AI quality concerns" would be possible, but might obscure a distinction between factual accuracy and pedagogical suitability.

The researcher also recognizes that Teacher B's concern is not simply insufficient training. The statement specifically concerns uncertainty about permissible data uploads and institutional governance.

These distinctions matter because the categories will influence how the study describes teachers' experiences and the institutional conditions surrounding AI adoption.

Step 4: Refine the Coding Framework

After examining additional transcripts, the researcher develops provisional category definitions and documents the decisions.

The framework remains provisional until the researcher examines its suitability across the relevant dataset. Additional categories may be required, and existing definitions may need refinement.

Step 5: Document What AI Contributed

The researcher retains the initial AI suggestions, relevant prompts, model information, category revisions, and rationale for the final decisions.

In the methodology, the researcher can explain that generative AI was used to suggest initial descriptive labels, while the researcher examined the transcripts, refined the framework, and retained responsibility for interpretation.

The resulting categories are not presented as independently discovered by AI. They are researcher-developed analytical decisions informed, in part, by AI-generated suggestions.

Misconceptions About Automatically Generated Coding Categories

When Should You Let AI Suggest Your Coding Categories?

AI-generated categories may be useful when researchers need preliminary ways to organize unfamiliar material or compare alternative interpretations. Their value depends on how the suggestions are incorporated into a methodologically defensible analytical process.

One useful strategy is to maintain a category-development log. Record each proposed category, its supporting passages, subsequent revisions, and the reasoning behind its acceptance or rejection.

This creates an auditable account of how the framework developed without pretending that every coding decision was predetermined or objectively discovered.

Researchers should also consider whether the model systematically favors familiar categories while overlooking culturally specific or less frequently expressed experiences. The possibility of systematic bias in AI coding is especially relevant when category generation influences what researchers subsequently look for in the data.

Finally, retain the prompts, model details, relevant outputs, and revisions needed to explain the procedure. These records contribute to documenting AI-assisted coding reproducibly, although they cannot eliminate the interpretive character of qualitative analysis.

Before Accepting AI-Generated Coding Categories

Frequently Asked Questions About AI-Generated Coding Categories

AI Can Suggest Categories, but Researchers Must Establish Their Meaning

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.