03 · What You Need to Know
How AI Develops Candidate Themes and Why Some Fail
What Makes a Theme Different From a Topic?
A topic identifies something participants discussed. A theme, particularly within reflexive thematic analysis, communicates a pattern of shared meaning organized around a central concept.
Suppose students repeatedly discuss internet connectivity, examination anxiety, and communication with instructors. AI might produce three themes using those exact labels.
These headings may be useful for organizing the data, but they mainly identify subjects. They do not necessarily explain what the accounts reveal about students' experiences.
A more interpretive candidate theme might be "Academic Participation Depends on Conditions Students Cannot Fully Control." Such a theme proposes a relationship among technological access, institutional expectations, and students' capacity to participate.
Whether this interpretation is justified depends on the complete dataset. The theme cannot be validated by its academic-sounding wording alone.
Topic summary
Identifies a subject that appears in the interviews, such as internet connectivity or examination anxiety.
Analytical theme
Communicates a meaningful pattern relevant to the research question, supported by interpretation of the data.
This distinction is especially important in thematic analysis, although other qualitative approaches may legitimately organize findings through descriptive categories rather than interpretive themes.
How Does AI Generate Themes From Transcripts?
A generative AI model can examine supplied text, identify semantic similarities, summarize repeated ideas, and propose relationships among passages. Depending on the instructions, it may generate themes directly from transcripts or use previously developed codes as an intermediate analytical structure.
For example, a model might receive the following initial codes:
- Avoiding camera use because of household visibility
- Fear of being judged during online discussions
- Unstable connectivity interpreted as disengagement
- Preference for text-based participation
It could group these under a broad category such as "Barriers to Online Participation."
Alternatively, it might propose that students experience participation as a negotiation between institutional expectations and their personal circumstances.
The second formulation makes a more explicit interpretive claim, but it is not automatically superior. The researcher must determine whether the claim is supported by the interviews and whether it addresses the research question.
The quality of the proposed themes also depends on the quality of the underlying initial qualitative coding. If codes misrepresent participants' accounts, grouping those codes into themes may reproduce or amplify the original errors.
Can AI Generate Themes Without First Coding the Interviews?
Yes. A researcher can ask AI to identify candidate themes directly from transcripts without providing a separate coding framework.
However, the absence of an explicit coding stage makes it more difficult to examine how the model moved from individual statements to broader interpretations.
A system may produce a theme such as "Students Demonstrate Resilience Despite Technological Challenges," but the researcher may not know which statements informed that conclusion or whether contrary accounts were considered.
Requiring the model to identify relevant transcript passages can improve traceability. It does not guarantee that the proposed theme is conceptually coherent or that the model considered all relevant evidence.
In some qualitative approaches, coding and theme development are closely interconnected rather than strictly sequential. The important issue is not whether the researcher followed a mechanical sequence, but whether the analytical decisions can be explained and defended.
What Should a Well-Developed Theme Contain?
Within an interpretive thematic analysis, a well-developed theme should offer more than a collection of related statements. It should communicate a meaningful pattern that contributes to answering the research question.
Several considerations help evaluate candidate themes:
-
Central organizing concept: The theme expresses a coherent analytical idea rather than an assortment of loosely related topics.
-
Connection to the research question: The pattern contributes to understanding the phenomenon being investigated.
-
Evidentiary grounding: The interpretation can be supported through authentic extracts examined in their original context.
-
Analytical coherence: The included material supports a meaningful relationship without requiring every participant to express an identical experience.
-
Distinct contribution: The theme adds something useful to the overall analysis rather than unnecessarily repeating another theme.
-
Appropriate interpretation: Claims do not exceed what the data and methodological framework can reasonably support.
These are considerations for analytical judgment, not a scoring formula. Different qualitative approaches may emphasize different criteria.
Why AI Often Produces Generic Themes
Generative AI may produce broad headings such as "Challenges," "Coping Strategies," "Benefits," and "Recommendations" because these categories offer a convenient way to organize many kinds of interview material.
Such categories are not necessarily incorrect. They may be appropriate for some descriptive research purposes. However, they can become analytically weak when the study seeks to explain how participants construct meaning or experience a phenomenon.
For instance, grouping accounts of financial difficulty, limited connectivity, and family obligations under "Challenges" may obscure the relationship among these experiences.
A more developed interpretation might investigate how students' educational participation depends on access to resources that institutions sometimes assume are universally available.
That interpretation must still be tested against the data. The objective is not to replace simple headings with elaborate ones, but to identify what the interviews actually permit the researcher to understand.
Can AI Recognize Themes That Are Not Frequently Mentioned?
AI can be prompted to examine less common accounts, but its summaries may privilege recurring or highly explicit statements.
Frequency is not universally equivalent to analytical importance. A single account may reveal a condition that complicates an emerging interpretation or identifies an experience overlooked by most participants.
Suppose 15 interviewees describe recorded lectures as convenient, while one participant explains that recordings are inaccessible because the household shares one device.
A theme emphasizing flexibility may reflect the dominant account but fail to explain the conditions under which flexibility is possible.
Examining contradictory and deviant cases can help researchers refine their interpretation rather than treating minority experiences as irrelevant exceptions.
Can AI Compare Themes Across Participant Groups?
Generative AI can organize candidate themes across participant groups when transcripts contain reliable group identifiers and sufficient contextual information.
For example, a researcher might compare how first-year and graduating students discuss online learning, or how students with different access arrangements experience synchronous classes.
However, a difference in how frequently a topic appears does not automatically establish a meaningful difference between groups. Interview questions, sample composition, transcript length, and participants' opportunities to discuss the topic may affect the comparison.
Researchers should distinguish differences in the available accounts from broader claims about populations. More detailed procedures for comparing themes across participant groups require attention to within-group variation and the study's sampling logic.
Why Theme Development Requires Contextual Interpretation
Interview statements are produced within conversations. Their meanings may depend on preceding questions, the participant's relationship with the interviewer, cultural conventions, and the broader narrative.
Consider the statement, "I guess we're used to it."
AI might interpret this as adaptation or resilience. Yet the participant could be expressing resignation, frustration, irony, or genuine acceptance.
Without examining the surrounding conversation, the researcher cannot confidently determine which reading is warranted.
Theme development therefore requires more than identifying semantic similarity. It involves evaluating contextual and culturally situated meanings that may not be apparent from isolated excerpts.
What Does Empirical Research Tell Us?
Published research has examined whether large language models can produce themes resembling those developed through conventional qualitative analysis.
Comparative studies in healthcare have reported meaningful overlap between AI-generated and human-generated thematic outputs under particular evaluation procedures. Other research has identified weaknesses involving contextual interpretation, inductive analysis, and the selection of supporting quotations.
These findings suggest that AI can contribute to candidate theme generation, but they do not establish that all themes produced by a model are analytically defensible.
Agreement between AI and human analysts is also not a universal criterion of thematic quality. In reflexive thematic analysis, different interpretations may be defensible when they are coherent, grounded in the data, and consistent with the researcher's theoretical orientation.
The relevant question is therefore not simply whether AI produces the same theme names as a human researcher, but whether the resulting interpretation can be justified.
Can AI Replace the Researcher in Developing Themes?
The answer depends partly on the methodological approach and what is meant by replacement.
In a structured analysis, AI may perform substantial classification and organizational work under a researcher-defined procedure. In reflexive thematic analysis, theme development is understood as an active interpretive process shaped by the researcher's theoretical assumptions and engagement with the data.
A model can generate a thematic narrative, but it does not thereby acquire the researcher's situated relationship to the participants or responsibility for the knowledge claims.
Delegating the final interpretation may therefore create methodological tensions, particularly when the study explicitly relies on researcher reflexivity.
The researcher should determine which activities can be assisted by AI without undermining the analytical commitments of the study.
04 · A Practical Example
Turning Interview Patterns Into a Candidate Analytical Theme
Hypothetical Example
How University Students Experience Camera Requirements
A researcher interviews students about participation in synchronous online classes. The research question asks how students experience expectations concerning camera use and visibility.
The following excerpts are invented for illustration:
-
Participant A: "I don't want my classmates to see how crowded our house is."
-
Participant B: "Our instructor thinks we're not paying attention when our cameras are off."
-
Participant C: "I turn off my camera because the connection becomes unstable when video is on."
-
Participant D: "I feel more connected when I can see everyone."
-
Participant E: "Sometimes I open my camera even when I'm uncomfortable because I don't want to be marked absent."
Stage One: AI Identifies Recurring Topics
An AI system might initially suggest:
- Privacy concerns
- Instructor expectations
- Internet connectivity
- Social connection
- Attendance requirements
These labels describe relevant subjects but do not yet explain how the experiences are connected.
Stage Two: AI Suggests a Candidate Theme
After being asked to examine relationships among the accounts, AI might propose:
Candidate Theme
Being Visible Becomes a Measure of Being Present
Camera use may operate as an institutional signal of participation, even when students' decisions to become visible are shaped by privacy concerns, technological limitations, and discomfort. The expectation to demonstrate presence visually may therefore create tension between institutional definitions of engagement and students' circumstances.
This candidate theme offers a possible explanation of how camera requirements acquire meaning within online classes. However, it may overemphasize institutional monitoring if the wider dataset contains substantial evidence that students experience visibility primarily as social connection.
Stage Three: The Researcher Examines the Proposed Interpretation
1. Return to the complete transcripts.
Examine whether participants consistently connect camera use with being perceived as present or engaged.
2. Check the supporting evidence.
Verify that quotations are authentic and that the proposed interpretation reflects their original context.
3. Investigate contradictory accounts.
Examine whether positive experiences of visibility complicate the candidate theme or suggest another meaningful pattern.
4. Refine the central organizing concept.
Determine whether the theme is primarily about surveillance, participation expectations, social connection, or the negotiation of visibility.
5. Review the wider dataset.
Assess whether the proposed pattern extends beyond the illustrative statements and contributes to answering the research question.
6. Document the analytical decisions.
Record how the theme was proposed, revised, supported, or rejected, including the role of AI assistance.
The example demonstrates that theme development involves constructing and evaluating an interpretation, not simply grouping quotations beneath a heading.
06 · What This Means for You
How to Evaluate AI-Generated Themes Before Using Them
Use AI-generated themes as propositions to investigate rather than findings to accept. The analytical task is to determine whether the proposed pattern explains something meaningful about the data and whether the evidence supports that explanation.
A Practical Theme Evaluation Framework
If the proposed theme merely names a subject
Examine what participants communicate about that subject and whether a meaningful pattern can be developed.
If several proposed themes overlap
Compare their central organizing concepts and determine whether they represent distinct analytical claims.
If a theme seems supported by only a few extracts
Review the wider dataset and assess its analytical relevance rather than rejecting it solely because of frequency.
If a theme ignores contradictory accounts
Reconsider its boundaries, qualifications, or underlying interpretation.
If the theme depends on an inferred motive or cultural meaning
Check the original language, surrounding interview, and contextual evidence before accepting the interpretation.
Ask AI to Explain the Evidence Behind Each Candidate Theme
A useful prompt might be:
"Using the supplied anonymized interview excerpts and research question, propose candidate themes that represent meaningful patterns rather than merely recurring topics. For each candidate, explain its central organizing concept, identify exact supporting passages, describe alternative interpretations, and identify excerpts that complicate the proposed pattern. Do not invent quotations or present the themes as final findings."
This may make the suggestions easier to evaluate, although the model's explanations remain subject to verification.
Keep Theme Development Iterative
Researchers may need to rename themes, divide broad themes, combine overlapping interpretations, or abandon patterns that initially appeared promising.
These revisions are not necessarily signs of poor analysis. They may reflect deeper engagement with the data and a more precise understanding of the phenomenon.
Maintain analytical records that explain how candidate themes evolved. Analytical memos can help document interpretive questions, alternative readings, and decisions concerning thematic boundaries.
Report AI Assistance Transparently
If AI materially influenced theme generation, document the relevant system, its analytical tasks, the information supplied, and how the researcher evaluated or revised its suggestions.
Follow applicable ethics, confidentiality, institutional, and publisher requirements. The purpose of reporting AI use in qualitative analysis is to make the analytical process understandable rather than simply acknowledge that a tool was used.
