03 · What You Need to Know
How AI-Assisted Coding Interacts With Qualitative Epistemology
What Are Epistemological Assumptions in Qualitative Research?
Epistemology concerns the nature of knowledge, including how knowledge can be developed, what constitutes justification, and how researchers understand the relationship between themselves and what they study.
In qualitative research, epistemological assumptions influence how researchers treat participants' accounts, interpret evidence, and justify analytical conclusions.
For example, some approaches emphasize systematic description of participants' reported experiences. Others examine how meaning is constructed through social interaction, language, institutional practices, or the research encounter itself.
These differences affect what counts as an adequate interpretation and which procedures are methodologically appropriate.
Epistemology should also be distinguished from ontology, which concerns assumptions about the nature of reality, and methodology, which concerns the rationale and organization of the research approach. These concepts are related, but they are not interchangeable.
Why Coding Is Not Necessarily a Neutral Technical Operation
Coding involves identifying and labeling features of qualitative material. Even when codes are descriptive, decisions must be made about which passages matter, where meaningful segments begin and end, and what the labels represent.
Consider a participant who says:
"I don't open my camera because I share a room with my siblings."
One researcher might code this as "Shared Living Space." Another might emphasize "Managing Privacy During Online Classes." A third might examine how expectations concerning visibility intersect with domestic circumstances.
These possibilities are not necessarily competing answers to a single classification problem. They may reflect different research questions and analytical orientations.
Generative AI also makes selections when producing codes. Its suggestions are influenced by the supplied instructions, available context, and patterns learned during training.
The concern is therefore not simply whether AI can generate initial qualitative codes, but how those generated labels influence what the researcher subsequently recognizes as meaningful.
Does Using AI Mean the Study Becomes Positivist?
No. Using a computational tool does not, by itself, establish a positivist epistemology.
Positivism is associated with particular assumptions about knowledge and scientific inquiry. It should not be reduced to the use of numbers, software, or standardized procedures.
Similarly, interpretivism does not necessarily prohibit all systematic organization or technological assistance.
What matters is how the procedures are understood and justified.
For example, a researcher might use AI to organize material while continuing to recognize that coding decisions are interpretive and context-dependent. Another researcher might treat AI-generated labels as objectively correct representations of meaning and use agreement statistics as the sole criterion of validity.
These practices involve different assumptions about coding and knowledge, even if both researchers use the same software.
Methodological congruence therefore cannot be established simply by identifying the technology employed.
How Do Different Epistemological Orientations Affect AI Coding?
Qualitative approaches vary considerably, and broad paradigm labels do not determine every analytical procedure. Nevertheless, examining their assumptions can help identify potential tensions.
| Orientation |
Relevant understanding of knowledge |
Questions for AI-assisted coding |
| Postpositivist approaches |
Knowledge claims may be evaluated through systematic procedures while recognizing uncertainty and limitations. |
Can coding rules be specified, tested, and evaluated without overstating objectivity? |
| Interpretivist approaches |
Understanding depends on participants' meanings and their social and contextual circumstances. |
Does AI preserve contextual meaning, or does it impose generalized categories? |
| Constructivist approaches |
Knowledge is understood as developed through interpretive and social processes rather than simply extracted from data. |
Whose interpretations shape the codes, and how is the researcher's role represented? |
| Critical approaches |
Knowledge production is examined in relation to power, inequality, ideology, and social structures. |
Whose perspectives are privileged by the model, and what power relations shape the analysis? |
These descriptions are necessarily simplified. A particular study may combine philosophical commitments that do not fit neatly into one category.
The table is intended to identify methodological questions, not to declare that AI is universally acceptable or unacceptable within any broad paradigm.
Why Treating Codes as Fixed Objects Can Create Epistemological Tension
Some AI coding workflows assume that every excerpt has a correct label that can be identified through sufficiently accurate classification.
This assumption may be appropriate for a narrowly defined deductive task, such as identifying whether participants explicitly report connectivity problems under a predefined codebook.
However, it becomes problematic when applied to interpretive work in which coding is understood as an active process of meaning-making.
For example, the same passage may support different analytical readings depending on the research question, theoretical framework, and surrounding material.
In such circumstances, disagreement between two interpretations does not automatically mean that one is incorrect.
Researchers should distinguish tasks involving classification according to agreed rules from tasks involving the development of interpretive meaning.
Can Coding Agreement Establish Interpretive Validity?
Agreement statistics can be useful in qualitative approaches that explicitly seek consistent application of defined categories.
For example, a research team might assess whether AI and human coders consistently identify statements meeting a specified code definition.
However, agreement does not universally establish the quality of qualitative interpretation.
Braun and Clarke distinguish coding reliability approaches from reflexive thematic analysis. In reflexive thematic analysis, different researchers may develop different, defensible interpretations of the same data, and coding agreement is not treated as the principal criterion of analytical quality.
Therefore, a high agreement coefficient between AI and human coding does not establish that the analysis is epistemologically consistent with reflexive thematic analysis.
Likewise, low agreement does not automatically invalidate an interpretive analysis when the methodology does not conceptualize coding as a search for one correct classification.
Does Applying a Predefined Codebook Change the Epistemology?
Not necessarily. Codebooks are compatible with several qualitative approaches, including forms of qualitative content analysis and framework analysis.
A codebook may provide definitions, inclusion criteria, exclusion rules, and examples that guide systematic classification.
AI can potentially assist with applying a predefined qualitative codebook, provided that the coding procedure is validated and consistent with the study's methodological commitments.
The epistemological issue arises when the researcher assumes that codebook application captures the complete meaning of participants' accounts or treats classification consistency as sufficient evidence of interpretive quality.
A structured coding framework may support the research question without exhausting the analytical possibilities contained in the data.
Why AI Can Influence What Researchers Notice
AI-generated codes may shape the researcher's subsequent attention, even when the researcher intends to treat them as provisional.
Suppose AI labels several excerpts as "Student Resilience." The researcher may begin interpreting related statements through that concept, paying less attention to expressions of resignation, frustration, or constrained choice.
This creates a risk of interpretive anchoring. The first generated categories can influence which patterns appear salient during later analysis.
The concern is not unique to AI. Human researchers also develop expectations and interpretive commitments. However, AI can introduce a ready-made conceptual structure before the researcher has sufficiently engaged with the material.
Researchers should examine how generated labels influence subsequent coding, theme development, and the treatment of contradictory evidence.
Can AI Change the Meaning of Inductive Analysis?
Inductive analysis is often described as developing analytical insights through engagement with the data rather than applying an entirely predetermined framework.
Asking AI to generate codes without supplying a codebook may appear inductive.
However, the model's suggestions are shaped by prior training and instructions. They do not emerge from the dataset independently of conceptual influences.
This does not automatically make the workflow deductive. Human inductive analysis is also informed by researchers' prior knowledge and theoretical sensibilities.
The more important question is whether the analytical process remains open to meanings that challenge initial expectations and whether the researcher can explain how the codes were developed.
AI-assisted inductive coding should therefore be described accurately rather than presented as entirely free of prior assumptions.
Why Reflexive Thematic Analysis Raises a More Specific Objection
Reflexive thematic analysis is particularly relevant because it conceptualizes coding and theme development as researcher-generated interpretive work.
In December 2025, Jowsey, Braun, Clarke, and numerous other qualitative researchers published a statement rejecting generative AI use in reflexive qualitative research. The authors argue that these approaches require a subjective, positioned, and reflexive human researcher and that generative AI is not methodologically congruent with those commitments.
This position is stronger than the general recommendation to verify AI-generated outputs.
Under that methodological view, human verification does not automatically resolve the problem of delegating interpretation to a system that does not occupy the researcher's situated role.
Researchers should therefore distinguish broad discussions of AI-assisted qualitative coding from the specific question of delegating interpretation in reflexive thematic analysis.
How Do Reflexivity and Positionality Enter the Epistemological Question?
In interpretive qualitative traditions, researchers may examine how their assumptions, experiences, relationships, and institutional positions shape the knowledge produced.
Generative AI can generate text about assumptions and positionality, but such text does not establish that the system possesses the researcher's lived or institutional relationship to the study.
Similarly, asking AI to explain its own coding decisions does not necessarily reveal the computational processes responsible for its output.
The researcher must therefore distinguish between generated explanations and genuine methodological accountability.
Questions about AI and reflexivity are related to epistemology because they concern how interpretive knowledge is developed and justified.
Can AI Introduce Assumptions From Its Training Data?
Yes. Generative AI outputs reflect patterns learned from training data and are shaped by model development and prompting.
These influences may affect how the system interprets language, assigns categories, and prioritizes explanations.
For example, a model may interpret reluctance to speak during online classes as low confidence while overlooking institutional expectations, linguistic practices, or cultural meanings.
Such interpretations can introduce assumptions that were not explicitly adopted by the researcher or established by participants' accounts.
Researchers should examine whether AI-generated concepts reflect the data or reproduce familiar explanations that are insufficiently supported.
This is particularly consequential in critical qualitative research, where the analysis may seek to challenge dominant assumptions rather than reproduce them.
Does Human Oversight Preserve Epistemological Consistency?
Human oversight can improve accuracy and accountability, but it does not automatically establish methodological congruence.
Suppose AI generates the codes and themes while the researcher checks the quotations and approves the final narrative.
The resulting analysis may be more accurate than an unreviewed AI output. Nevertheless, the researcher has not necessarily undertaken the interpretive activities required by the declared methodology.
In contrast, a structured coding study might legitimately assign substantial classification work to AI if the research design permits it and the procedure is adequately validated.
The difference concerns the role of interpretation within the methodology, not merely the amount of human involvement.
Can Researchers Develop New AI-Integrated Qualitative Methodologies?
Researchers may propose and evaluate new analytical approaches that incorporate generative AI.
However, methodological innovation requires more than adding AI to an established procedure.
A defensible approach should explain its assumptions about knowledge, the relationship between human and machine-generated interpretations, the criteria used to evaluate analytical quality, and the limitations of the resulting claims.
It should also distinguish itself from established methodologies when its procedures differ in consequential ways.
Researchers should not assume that retaining a familiar methodological label is sufficient when the underlying analytical process has changed.
04 · A Practical Example
When AI Coding Conflicts With an Interpretivist Research Design
Hypothetical Example
Understanding Students' Experiences of Mandatory Camera Use
A researcher adopts an interpretivist approach to investigate how university students experience mandatory camera use during online classes.
The study seeks to understand how participants construct meaning around visibility, participation, privacy, and classroom expectations.
The researcher provides the following invented excerpts to an AI system:
-
Participant A: "I don't open my camera because I share a room with my siblings."
-
Participant B: "Our instructor thinks we're not listening if our cameras are off."
-
Participant C: "I feel more connected when everyone turns on their cameras."
-
Participant D: "Sometimes I turn my camera on even when I'm uncomfortable because I don't want to be marked absent."
What AI Might Produce
The model generates four codes:
- Privacy Concerns
- Perceived Instructor Monitoring
- Social Connection
- Compliance With Attendance Requirements
These labels may be reasonable initial suggestions. However, suppose the researcher treats them as objectively correct categories and instructs AI to assign every subsequent excerpt to exactly one of the four codes.
The procedure may begin to conflict with the study's interpretivist purpose because it restricts the analysis to a fixed classification structure before the researcher has adequately examined how participants construct meaning.
Where the Epistemological Tension Appears
| Research commitment |
AI coding practice |
Potential tension |
| Meaning is interpreted in context. |
Each isolated sentence receives a fixed label. |
Contextual meanings may be reduced to predefined categories. |
| Multiple readings may be defensible. |
One code is treated as the correct answer. |
Interpretive plurality may be mistaken for classification error. |
| The researcher actively develops understanding. |
AI determines the coding structure and thematic relationships. |
The researcher's interpretive role may be displaced. |
| Analytical quality requires methodological coherence. |
High AI-human agreement is treated as proof of validity. |
A reliability criterion may replace the intended interpretive standard. |
What a More Defensible Decision Might Involve
The researcher has two broad possibilities, depending on the intended research purpose.
If the study remains committed to a researcher-centered interpretive methodology, the analytical process should follow that methodology's requirements. In reflexive thematic analysis, this includes taking seriously the published objections to generative AI involvement.
If the researcher's actual objective is systematic classification of statements into predefined categories, a structured qualitative content analysis or another suitable approach may better describe the procedure.
Changing the label alone is not sufficient. The research question, analytical assumptions, procedures, and quality criteria must be aligned.
A Methodological Congruence Review
1. State the epistemological commitments.
Explain how the study understands participant meaning, researcher interpretation, and the justification of knowledge claims.
2. Identify the analytical role assigned to AI.
Distinguish data organization, predefined classification, initial coding, theme development, and substantive interpretation.
3. Examine the assumptions embedded in the workflow.
Determine whether the procedure treats codes as fixed categories, researcher-generated interpretations, or something else.
4. Evaluate the quality criteria.
Check whether agreement, reliability, reflexivity, theoretical coherence, or other standards are appropriate to the chosen approach.
5. Resolve methodological inconsistencies.
Revise the analytical procedure or reconsider the methodological framework when the two cannot be justified together.
6. Report the actual process.
Explain how the analysis was conducted, including material AI involvement and the researcher's analytical responsibilities.
The purpose of this review is not to declare that one technology determines the research paradigm. It is to ensure that the procedures used to produce findings remain consistent with the study's claims about knowledge.
06 · What This Means for You
How to Evaluate Methodological Congruence Before Using AI Coding
Before incorporating generative AI into qualitative coding, identify the assumptions that justify your analytical approach.
Ask what the codes are intended to represent, how they are developed, and what would count as evidence that the analysis is methodologically sound.
These questions are more consequential than whether AI can generate the coding table quickly or produce labels that resemble those developed by human analysts.
Match the AI Workflow to the Methodological Commitments
If the study uses predefined categories for structured classification
Consider whether AI-assisted coding can be validated using criteria appropriate to the analytical approach.
If the study emphasizes contextual interpretation
Examine whether AI-generated labels reduce meanings to generalized categories or introduce unsupported assumptions.
If the methodology requires researcher-led reflexive interpretation
Do not assume that AI-generated codes or themes are compatible merely because the researcher reviews the output.
If AI determines the substantive analytical framework
Reconsider whether the declared methodology accurately represents how the findings were produced.
If the analytical workflow is substantially modified
Explain the modification, justify its epistemological assumptions, and evaluate the resulting method on its own terms.
Use a Methodological Review Before Automating Coding
Where AI-assisted analysis is methodologically permissible, a researcher may use the following questions to examine a proposed workflow:
"Evaluate the proposed qualitative coding procedure against the study's stated epistemological assumptions and methodology. Identify where the workflow treats codes as fixed classifications, where it requires interpretive judgment, and whether its quality criteria are appropriate. Distinguish methodological inconsistencies from technical limitations. Do not assume that AI coding is inherently objective or that human review automatically resolves incompatibility."
An AI-generated methodological critique is not itself an authoritative determination of congruence. Researchers should verify its claims against the methodological literature and, where appropriate, seek specialist advice.
Document How AI Influences Knowledge Production
When AI is used, record which analytical tasks it performed, what material it processed, and how its outputs influenced coding and interpretation.
Consider whether generated labels shaped subsequent attention, whether alternative interpretations were examined, and whether the analytical framework changed during the process.
These records may support methodological transparency, but they should not be presented as evidence that the researcher undertook interpretive activities that were actually delegated to AI.
Keep Epistemological Justification Separate From Disclosure
Reporting that AI was used is necessary when required by applicable institutional or publication policies, but disclosure does not establish that the analytical procedure was appropriate.
Researchers must separately explain why the procedure fits the study's assumptions and how the resulting interpretations were evaluated.
Detailed reporting of AI use in qualitative analysis should accurately represent the analytical process without concealing methodological tensions.
Consider Ethical and Contextual Consequences
AI-assisted coding may involve processing sensitive interview transcripts, including personal experiences and indirectly identifying information.
Researchers should verify that the proposed use complies with ethics approval, participant consent, institutional requirements, data-use agreements, and applicable privacy laws.
They should also consider whether the model's linguistic and cultural assumptions affect the representation of participants' accounts, particularly when working with marginalized groups or multilingual data.
