03 · What You Need to Know
What Researchers Should Disclose About AI-Assisted Qualitative Analysis
Why Is AI Disclosure a Methodological Issue?
Qualitative methods sections explain how researchers moved from collected material to analytical findings. Readers need to understand the relationship between the data, the analytical procedures, and the interpretations presented.
When generative AI contributes to this process, its role may influence which statements receive attention, how excerpts are categorized, and which conceptual relationships are proposed.
For example, a system used only to correct transcription formatting has a different methodological role from one that generates the final themes of a study.
Both uses may warrant documentation, but the latter directly concerns the development of analytical knowledge.
Reporting should therefore reflect the actual significance of AI involvement rather than treating every application as equivalent.
What Information Should Be Reported?
There is no single universally mandated reporting checklist for all AI-assisted qualitative studies. However, several elements are important for understanding and evaluating the analytical process.
| Reporting element |
What to describe |
Why it matters |
| AI system |
Tool, provider, model or version when available, and relevant configuration |
Identifies the technology involved |
| Purpose of use |
Specific tasks assigned to AI |
Distinguishes technical assistance from analytical contribution |
| Data processed |
Type and scope of materials supplied, without disclosing confidential information |
Clarifies what evidence was available to the system |
| Instructions |
Relevant prompts, codebooks, coding rules, and analytical constraints |
Explains how outputs were elicited |
| Processing workflow |
Segmentation, batching, preprocessing, and relevant changes during analysis |
Identifies procedures that may affect consistency and completeness |
| Human involvement |
Who reviewed, revised, rejected, or developed analytical outputs |
Clarifies the distribution of analytical responsibility |
| Verification |
How quotations, coding decisions, and interpretations were checked |
Allows readers to evaluate quality-control procedures |
| Ethical safeguards |
Relevant permissions, confidentiality arrangements, and data-processing protections |
Addresses participant protection and responsible data handling |
| Methodological justification |
Why AI involvement was considered appropriate for the analytical approach |
Addresses compatibility with the research design |
Not every study needs an extensive discussion of every element. The level of detail should be proportionate to AI's actual contribution and the information necessary to evaluate the analysis.
Why Naming the AI Tool Is Not Enough
Statements such as "ChatGPT was used for thematic analysis" identify a tool but leave the analytical procedure largely unexplained.
Readers cannot determine whether the system summarized transcripts, generated initial codes, applied an existing codebook, developed candidate themes, or wrote the final findings.
These activities differ substantially in their methodological implications.
A more informative account specifies the analytical tasks and the researcher's involvement. For example, a researcher might explain that AI suggested provisional codes for selected excerpts, while the researcher examined the original transcripts and made the final coding decisions.
Such a description is appropriate only if it accurately represents what occurred.
Researchers should avoid using vague terms such as "assisted," "supported," or "facilitated" when those terms conceal substantial delegation of interpretation.
Should You Report the Exact Model and Version?
Where available, researchers should identify the AI system and relevant model information with sufficient precision to distinguish the technology used.
For example, reporting only "ChatGPT" may be insufficient when the particular model, deployment arrangement, or configuration materially affects the procedure.
Useful information may include the provider, model identifier, access method, approximate period of use, and relevant settings.
However, commercial systems can change without researchers receiving complete information about underlying updates. A model name does not necessarily identify an immutable computational system.
Researchers should report what they can verify and avoid inventing version information that was not recorded.
Where reproducibility is important, preserving the original outputs and relevant settings may be more informative than relying on a product name alone.
Do You Need to Include Every Prompt?
Not necessarily. The appropriate level of prompt documentation depends on whether the instructions materially influenced the analytical procedure.
If AI was used to apply a predefined codebook, the instructions defining code application, exclusions, overlapping codes, and uncertainty handling may be essential to understanding the analysis.
If AI generated themes, prompts specifying the research question, analytical orientation, and expected outputs may be methodologically consequential.
Researchers can summarize the procedure in the methods section and provide representative or complete relevant prompts in supplementary materials when permitted and useful.
Prompts containing confidential participant information should not be published without appropriate safeguards.
Researchers should also document substantive prompt revisions. A final prompt alone may conceal earlier iterations that influenced the analytical direction.
Should You Report Which Data Were Submitted to AI?
Yes, to the extent necessary to explain the procedure without compromising confidentiality.
Readers should understand whether AI processed complete transcripts, selected excerpts, anonymized summaries, coded passages, or another representation of the original data.
This distinction matters because the model's interpretation depends on what information it received.
For example, a system analyzing isolated quotations may lack the conversational context necessary to distinguish irony, resignation, or indirect disagreement.
Similarly, a model processing only researcher-selected excerpts cannot independently evaluate patterns across material that was never supplied.
Researchers should report important selection and preprocessing decisions rather than implying that AI examined the entire dataset when it processed only a subset.
Why Batching and Context Limits Matter
Large qualitative datasets may exceed the amount of information a model can process reliably in a single interaction.
Researchers may divide transcripts into batches, summarize individual interviews, or use retrieval systems to provide selected passages.
These procedures can influence the resulting analysis. Patterns spanning different batches may be overlooked, and earlier summaries may remove details that later become analytically important.
Where batching materially affects the workflow, researchers should explain how transcripts were divided, how outputs were consolidated, and how consistency was evaluated.
It is also important to distinguish technical processing limits from methodological sampling decisions.
Dividing transcripts because of system constraints is not equivalent to selecting participants according to a qualitative sampling strategy.
How Should AI-Assisted Coding Be Reported?
Reporting should distinguish whether AI generated new codes or applied an existing coding framework.
When AI generated initial qualitative codes, researchers should explain how suggestions were produced, evaluated, revised, or rejected.
When AI applied a predefined codebook, reporting should describe the coding rules and relevant validation procedures.
For structured coding, researchers may report agreement or classification performance when those measures are appropriate to the methodology.
However, coding agreement should not automatically be presented as evidence of quality in interpretive approaches that do not conceptualize coding as a search for one correct classification.
The methods section should describe the analytical logic actually used rather than adopting familiar reliability terminology without methodological justification.
How Should AI-Generated Themes Be Disclosed?
If AI proposed or developed themes, this contribution should be distinguished from researcher-generated theme development.
For example, a researcher might report that the system identified candidate patterns among coded extracts, which were subsequently examined and revised through researcher-led analysis.
Alternatively, if AI generated the substantive thematic structure and the researcher primarily approved it, the report should accurately describe that arrangement.
These are not methodologically equivalent workflows.
Researchers should also explain how supporting quotations were verified and how contradictory evidence was examined.
The ability to produce a polished thematic table does not establish that AI-generated themes represent defensible interpretations.
Why Human Verification Must Be Described Precisely
Statements such as "All AI outputs were checked by the researcher" may sound reassuring but provide little information about what checking involved.
Verification can refer to several different activities:
- Confirming that quotations appear in the original transcripts.
- Checking participant identifiers and excerpt attribution.
- Evaluating whether code assignments follow predefined definitions.
- Examining whether interpretations are supported by the wider dataset.
- Identifying contradictory cases and alternative explanations.
- Revising or rejecting AI-generated analytical claims.
These procedures address different potential errors. Checking quotations does not establish that the interpretation is methodologically appropriate, while evaluating a theme does not guarantee that every supporting quotation is accurate.
Researchers should describe which procedures were actually undertaken and avoid implying that every output received exhaustive review when only selected material was examined.
Should You Report AI-Generated Analytical Memos?
If AI materially contributed to analytical memoing, the researcher should explain the nature of that contribution.
For example, AI may have organized existing notes, proposed alternative interpretations, or generated questions concerning emerging patterns.
These uses differ from producing first-person accounts of researcher reflections or decisions that never occurred.
An AI-generated statement describing how the researcher supposedly reconsidered an assumption should not be presented as an authentic analytical record unless it accurately reflects the researcher's documented experience.
Reporting AI-assisted analytical memoing should therefore preserve the distinction between generated suggestions and genuine researcher reasoning.
How Should Methodological Compatibility Be Addressed?
Disclosure and methodological justification are related but distinct.
A researcher may transparently report that AI generated all codes and themes. That disclosure makes the procedure visible, but it does not establish that the procedure satisfies the declared qualitative methodology.
For example, Braun and Clarke's reflexive thematic analysis conceptualizes coding and theme development as researcher-generated interpretive activities.
In a 2025 statement, Jowsey, Braun, Clarke, and numerous other qualitative researchers explicitly rejected generative AI use in reflexive qualitative research, arguing that the technology is not methodologically congruent with approaches requiring a subjective, positioned, and reflexive human researcher.
This objection is particularly relevant when a study claims reflexive thematic analysis while delegating interpretation to AI.
Researchers should therefore explain not only what AI performed but also why the analytical procedure fits the study's epistemological and methodological commitments.
What Do Publisher and Research Reporting Policies Require?
Publication policies increasingly address generative AI use, although their scope and requirements vary.
The International Committee of Medical Journal Editors states that authors should disclose AI-assisted technologies used in submitted work, describe their use, and remain responsible for the accuracy and integrity of the manuscript. Its recommendations also distinguish appropriate reporting locations according to how the technology was used.
Nature Portfolio's editorial policies similarly emphasize author accountability and disclosure of relevant generative AI use, while recognizing that routine language editing may be treated differently from substantive content generation.
These policies should not be treated as identical or assumed to govern every journal. Researchers must consult the requirements applicable to their intended publication venue.
General AI disclosure policies also do not replace methodological reporting. A journal may require acknowledgment of AI assistance without specifying every detail necessary to evaluate an AI-assisted qualitative analysis.
Where Should AI Use Be Reported in a Research Article?
The appropriate location depends on the nature of AI involvement and the journal's requirements.
| Manuscript location |
Information that may belong there |
| Methods |
Analytical tasks, AI system, processing workflow, coding procedures, verification, and methodological justification |
| Ethics or data governance statement |
Relevant permissions, confidentiality safeguards, and approved processing arrangements |
| Limitations |
Consequential uncertainty, model errors, incomplete context, and limitations of the analytical procedure |
| AI-use declaration |
Concise disclosure required by the publisher or institution |
| Supplementary materials |
Relevant prompts, coding instructions, validation details, and nonconfidential analytical records |
These locations can serve different purposes. A brief declaration may identify AI involvement, while the methods section explains how that involvement affected the production of findings.
How Should Participant Confidentiality Be Reported?
Qualitative transcripts may contain sensitive information, including indirect identifiers that remain recognizable after names are removed.
Researchers should report relevant data protection arrangements without revealing confidential information or security-sensitive details.
Depending on the study, this may include whether transcripts were de-identified, whether an institutionally approved environment was used, and whether the proposed processing was covered by ethics approval and participant consent.
However, researchers should not claim that data were anonymous merely because names were replaced with participant codes.
Likewise, the use of a commercial AI service should not automatically be described as confidential or secure without examining the applicable service terms and processing arrangements.
Where the original research protocol did not authorize the proposed AI processing, researchers should seek appropriate institutional guidance rather than assume that retrospective disclosure resolves the issue.
Does Reporting Prompts Make the Analysis Reproducible?
Prompt documentation can improve procedural transparency, but it does not guarantee identical results.
Generative AI systems may produce different outputs across repeated interactions. Changes in models, settings, system instructions, or available context can also affect responses.
Even when prompts are preserved, another researcher may not be able to reproduce the same analytical output exactly.
Researchers should distinguish reproducibility of the documented procedure from exact reproduction of every generated sentence.
For some studies, preserving coding outputs, researcher decisions, model information, and relevant analytical records may be more useful than relying on prompts alone.
What Should You Do if AI Use Was Not Documented During Analysis?
Researchers should reconstruct only what can be established from authentic records.
Relevant evidence may include saved prompts, chat histories, exported outputs, dated analytical notes, codebook versions, and manuscript revisions.
If the precise model version or sequence of prompts cannot be recovered, the researcher should acknowledge the limitation rather than invent details.
Similarly, a researcher should not claim to have independently developed themes if the available records show that the substantive thematic structure originated from AI.
Retrospective documentation can improve transparency, but it cannot transform activities that did not occur into an authentic analytical history.
04 · A Practical Example
Writing an AI-Use Disclosure for Qualitative Interview Analysis
Hypothetical Example
A Study of Students' Experiences of Online Learning
A researcher conducts 20 semi-structured interviews concerning university students' experiences of participation in synchronous online classes.
The study uses a structured qualitative content analysis approach. The researcher develops an initial coding framework, then uses an institutionally approved generative AI environment to suggest code assignments for de-identified transcript excerpts.
The researcher compares the assignments with the codebook, examines ambiguous cases, reviews the original transcripts, and develops the final interpretation.
The following disclosures illustrate different levels of methodological detail. They are examples, not statements about an actual completed study.
Example One: A Disclosure That Is Too Vague
Insufficient Methodological Detail
"ChatGPT was used to assist with the qualitative data analysis. All results were verified by the researcher."
This statement does not explain which analytical tasks were performed, what verification involved, or whether AI influenced the final interpretations.
Example Two: A More Informative Methods Description
Illustrative Methods Paragraph
"Generative AI was used to support the application of a researcher-developed coding framework to de-identified interview excerpts. The system received the code definitions, inclusion and exclusion criteria, and instructions to identify supporting passages for each proposed assignment. The researcher reviewed the AI-generated coding suggestions against the original transcripts, revised or rejected assignments that did not satisfy the codebook, and examined ambiguous and contradictory cases. The final analytical categories and interpretations were developed by the researcher. The AI system, model information, processing arrangements, and relevant prompts were documented as part of the analytical record."
This paragraph describes the division of analytical work more clearly. In an actual manuscript, it should be supplemented with the specific system, relevant processing details, and validation information appropriate to the study.
It must not be copied unchanged if the researcher did not perform the activities described.
Example Three: When AI Actually Generated the Themes
Suppose the workflow was different. The researcher uploaded transcripts, instructed AI to generate the themes, and then reviewed the resulting thematic narrative.
Disclosure of Substantive AI Involvement
"A generative AI system was used to propose the initial thematic structure from the interview transcripts. The system generated candidate themes and supporting excerpts, which the researcher subsequently examined against the original data. The researcher reviewed quotation accuracy, considered alternative interpretations, and revised the generated thematic structure. Because AI contributed directly to theme development, the analytical procedure is described as AI-assisted rather than exclusively researcher-generated."
This description is more transparent about the model's role. However, it does not automatically establish that the approach was methodologically appropriate.
If the study claims reflexive thematic analysis, substantive AI-generated theme development raises a methodological objection that cannot be resolved through disclosure alone.
Example Four: A Concise AI-Use Declaration
Illustrative Declaration
"Generative AI was used during qualitative data analysis for the tasks described in the Methods section. The authors evaluated the generated outputs against the original research material and retain responsibility for the accuracy, interpretation, and reporting of the findings."
This declaration may complement a detailed methods description. It should not replace the explanation of the analytical procedure when AI materially influenced the findings.
A Practical Reporting Workflow
1. Identify every material AI contribution.
Distinguish transcription, summarization, coding, theme development, interpretation, and manuscript drafting.
2. Establish the actual processing procedure.
Record what material was supplied, which system was used, and how instructions and outputs were managed.
3. Describe human analytical involvement.
Explain who developed the coding framework, evaluated suggestions, revised interpretations, and approved the final claims.
4. Document verification.
Identify the checks performed on quotations, coding decisions, contextual meanings, and contradictory evidence.
5. Assess methodological compatibility.
Determine whether the AI-assisted procedure fits the study's analytical and epistemological assumptions.
6. Address ethics and confidentiality.
Report relevant permissions and data protection arrangements without exposing participant information.
7. Follow publication requirements.
Include the necessary methods description, AI declaration, limitations, and supplementary information.
The objective is to make the analytical procedure understandable without creating the impression that the researcher performed work that was actually delegated to AI.
06 · What This Means for You
How to Write an Accurate AI-Use Statement for Your Qualitative Study
Begin by describing the analytical process as it actually occurred. Do not start with a generic disclosure template and then adjust the history of the research to fit it.
The amount of detail should reflect the significance of AI involvement. A minor formatting task may require relatively little explanation, while AI-generated coding or theme development requires a more substantial methodological account.
Match the Disclosure to the AI Contribution
If AI was used only for technical formatting
Describe the technical task where required and avoid implying that AI performed substantive analysis.
If AI suggested initial codes
Explain how suggestions were generated, evaluated, revised, and incorporated into the coding process.
If AI applied a predefined codebook
Report the coding framework, relevant instructions, validation procedures, and handling of uncertain assignments.
If AI developed candidate themes
Identify the model's contribution to theme development and distinguish it from the researcher's subsequent analytical work.
If AI produced the substantive interpretation
Disclose the delegation explicitly and evaluate whether the procedure is compatible with the claimed methodology.
If AI-use records are incomplete
Report only what can be established and acknowledge consequential gaps rather than inventing methodological details.
Use a Reporting Template Based on Actual Research Records
The following structure can help organize a methods description:
"During qualitative data analysis, [AI system and model, where known] was used for [specific analytical tasks]. The system processed [type and scope of permitted data] using [relevant prompts, coding instructions, or analytical framework]. AI-generated outputs were evaluated through [actual verification procedures]. The researcher was responsible for [specific analytical decisions]. Relevant data protection arrangements included [verified safeguards]. The role of AI and limitations of the procedure were considered in relation to [qualitative methodology]."
Replace each field with verifiable information. Do not include procedures, safeguards, or researcher activities that were not actually undertaken.
Separate AI Disclosure From Methodological Justification
A disclosure explains what happened. Methodological justification explains why the procedure is appropriate for the research question and analytical approach.
Both may be necessary when AI materially influences qualitative interpretation.
For example, a study using deductive qualitative content analysis might justify AI-assisted codebook application through clearly defined categories and documented validation. A study claiming reflexive thematic analysis faces different methodological questions when substantive interpretation is delegated.
The issue is not whether the researcher used a familiar tool. It is whether the actual procedure supports the type of knowledge the study claims to produce.
Preserve the Records Needed to Support the Disclosure
Where feasible and ethically permitted, retain relevant prompts, model information, coding instructions, AI outputs, researcher corrections, analytical memos, and records of theme revisions.
These materials can support accurate reporting and help researchers explain how interpretations developed.
However, maintaining records does not require publicly releasing confidential transcripts or identifiable participant information.
Data-sharing decisions should remain consistent with consent, ethics approval, institutional policy, and applicable legal requirements.
Check the Journal's Current Requirements Before Submission
Publisher policies differ and may change. Before submitting a manuscript, consult the target journal's current instructions concerning AI disclosure, authorship, research methods, and supplementary materials.
Do not assume that compliance with one publisher's declaration format automatically satisfies another journal's requirements.
Where uncertainty remains, the journal's editorial office or the relevant institutional research authority may provide guidance.
