How to Report AI-Assisted Thematic Analysis in Your Methods Section

If you used AI to help code your data or draft candidate themes, say so in your Method section, name the tool and version, describe exactly what it did, and explain how you checked its output. Every major citation style now has a position on how to disclose and cite generative AI, and Journal Article Reporting Standards for Qualitative Research (JARS-Qual) already expect enough procedural detail for another researcher to follow your analysis (Levitt et al., 2018). Leaving AI assistance out is not a shortcut. It is a gap a reviewer will notice.

This guide sets out exactly what to write, where it goes in your manuscript, and how to phrase it so it reads as careful methodology rather than an apology. For more background on why AI is gaining traction in academic research and whether its use is truly valid, read our comprehensive overview on AI thematic analysis and its academic validity.

Why disclosure is not optional anymore

APA Journals updated its policy on generative AI in 2024: when a large language model is used in preparing a manuscript, that use must be disclosed in the Method section and cited, the same way you would cite any other software (McAdoo, Denneny and Lee, 2025). AI cannot appear as an author, but its role in your analysis has to be visible.

This is not just an APA rule. A 2025 position statement from Cochrane, the Campbell Collaboration, JBI, and the Collaboration for Environmental Evidence, four of the bodies that set standards for rigorous evidence synthesis, sets a similar bar for qualitative work: authors should declare AI use whenever it makes or suggests a judgment call, including coding, categorizing, or extracting meaning from data, and should name the tool, the version, the date used, and exactly which part of the process it touched. The same statement adds a useful distinction: AI used only to check spelling or grammar generally does not need disclosure, but AI used to interpret, code, or synthesize your data does.

Reviewers who study thematic analysis specifically already have a language for the underlying concern. Braun and Clarke's (2019) description of theme development as a reflexive process, shaped by the researcher's judgment rather than a neutral extraction from the data, is exactly what makes an undisclosed AI pass in your coding worrying to a supervisor. If your themes were partly generated by a model instead of built entirely through your own reflexive engagement with the data, the reader needs that information to judge your findings fairly.

Where this goes in your manuscript

For most qualitative papers, AI disclosure belongs in the Method section, typically in a subsection describing your data analysis procedure, sitting alongside your account of the six phases of thematic analysis. If your paper is a literature review or another format without a formal Method section, APA guidance places the disclosure in your introduction instead (McAdoo, Denneny and Lee, 2025).

Write it as a normal part of your procedure, not as a caveat tucked into a footnote or limitations paragraph. A sentence like this works:

“Initial coding of the interview transcripts was supported by [tool name and version] (Company, Year), a large language model-based tool, which generated a first-pass codebook from the raw transcripts. Each AI-suggested code was reviewed against the original data extract by the first author, who accepted, modified, or discarded the suggestion before proceeding to theme development in accordance with Braun and Clarke's (2006) six-phase framework.”

That single paragraph does four things a reviewer will look for: names the tool, states which phase it touched, states that a human reviewed every output, and ties the process back to a recognized methodological framework.

What to include, point by point

Drawing on the disclosure elements set out in the Cochrane/Campbell/JBI/CEE position statement, adapted here for thematic analysis specifically, a complete disclosure covers:

  • The tool, version, and date used. Name the specific product and model version, not just "AI" or "ChatGPT" generically. Models change quickly, and a vague reference makes your procedure impossible to replicate. If the tool is version-locked or updates automatically, note the date you used it.
  • Which phase of the analysis it supported. Be precise. "AI assisted with the analysis" tells a reviewer nothing. "AI assisted with Phase 2 (generating initial codes)" tells them exactly what to scrutinize and what remained entirely your own interpretive work.
  • What the human researcher did with the output. This is the part that determines whether your use of AI reads as rigorous or careless. State plainly that you reviewed, edited, or overrode AI suggestions, and roughly how. If you kept a record of changes, mention that a log exists and where it can be found, such as a supplementary file or appendix.
  • Why the tool was used. A short rationale, usually efficiency on a large dataset or consistency across a big batch of transcripts, helps a reviewer understand this was a deliberate methodological choice rather than something you're only mentioning because a policy made you.
  • Any prompts or instructions given to the tool, if they materially shaped the output. APA's own guidance on citing generative AI recommends including the prompt text when it is relevant to interpreting the output, sometimes as an appendix if lengthy (McAdoo, Denneny and Lee, 2025).

Citing the AI tool correctly, in your style

Every major style treats the AI company, not the model itself, as the closest thing to an author, since none of them recognize an AI tool as capable of authorship. Beyond that shared starting point, the four styles diverge in ways worth getting right, because a citation in the wrong convention is its own small credibility problem in a methods section meant to demonstrate rigor.

APA (7th edition)

Treats the AI's output as software output. The company is the author, the year is the model's release year, and the tool name with its version goes in the title position, followed by a bracketed description (McAdoo, Denneny and Lee, 2025):

Reference list: OpenAI. (2024). ChatGPT (GPT-4) [Large language model]. https://chat.openai.com/chat
In-text: (OpenAI, 2024)

Harvard

Harvard is not one fixed standard; it's a family of author-date conventions, and UK universities in particular each publish their own house variant, so check your institution's guide first. The widely used Cite Them Right interpretation follows a pattern close to APA: creator, year in round brackets, italicized title, a bracketed medium descriptor, then the access details (Ulster University Library, n.d.; UCD Library, 2025):

Reference list: OpenAI (2024) ChatGPT (GPT-4) [Large language model]. Available at: https://chat.openai.com/chat (Accessed: 14 March 2026).
In-text: (OpenAI, 2024)

MLA (9th edition)

The most structurally different of the four. MLA does not use the model as a title the way APA and Harvard do; instead, your prompt becomes the title of the source, in quotation marks, followed by the tool name in italics as the container (MLA Style Center, 2025):

Works Cited: “Generate an initial codebook from the following interview transcript excerpts” prompt. ChatGPT, version GPT-4, OpenAI, 14 Mar. 2026, chat.openai.com.
In-text: (“Generate an initial codebook...”)

Chicago

Chicago's own guidance is the outlier here: it explicitly does not require a bibliography or reference list entry for AI-generated content in most cases, treating an AI chat like a personal communication, since your specific conversation typically isn't retrievable by another reader (Chicago Manual of Style, n.d.). Instead, disclose the tool in-text or in a numbered footnote:

Footnote: 1. Text generated by ChatGPT, OpenAI, 14 March 2026, https://chat.openai.com/chat.
Author-date in-text: (ChatGPT, 2026)

A general rule across all four: whichever style you use, name the specific model and version, not just the brand ("ChatGPT" alone is not enough; "ChatGPT, GPT-4" or "Claude, Sonnet 4" is), and record the date you actually used it. Models update silently and often, and a vague citation makes your procedure impossible for another researcher to verify or replicate, which is the entire point of citing a tool in a methods section in the first place.

A worked example

Here is a fuller version of a data analysis subsection that reports AI use correctly, adapted to a hypothetical study of 22 interview transcripts:

“Data were analyzed using reflexive thematic analysis following the six-phase process described by Braun and Clarke (2006). After familiarization with the full dataset (Phase 1), initial coding (Phase 2) was supported by [Tool Name, Version] (Company, Year), a large language model-based coding assistant, which generated a first-pass codebook of 118 codes from the 22 transcripts. The first author reviewed every AI-suggested code against its source extract, discarding 31 codes judged too literal or unsupported by the surrounding context, and modifying the wording of 44 others to better reflect the researcher's interpretation. This process is documented in a coding log available as Supplementary Material. Theme development (Phases 3 through 5) was conducted manually by the research team through iterative review of the revised codebook, without further AI involvement.”

Notice what this example does not do. It does not claim the AI generated the final themes. It does not treat the tool's output as final without a stated review process. And it gives a reviewer a specific number they could, in principle, check against your supplementary materials.

Common mistakes reviewers flag

  • Vague language that hides the extent of AI involvement. Phrases like “AI tools were used to assist with data analysis” without specifying what, where, or how much invite suspicion rather than reassurance. Reviewers read vagueness here as evasiveness, even when the underlying use was entirely reasonable.
  • No mention of human review. If your write-up describes what the AI did but never states that a person checked its work, a careful reader will assume no one did.
  • Treating disclosure as a limitations-section afterthought. AI use is part of your method, not a weakness to confess at the end. Reporting it in the Method section, in the same factual tone you'd use to describe your sampling strategy, signals that it was a considered methodological decision.
  • Citing the tool without describing the process. A reference list entry alone satisfies a citation requirement but not a methodological one. JARS-Qual is concerned with whether another researcher could follow your procedure, not just whether you named your software (Levitt et al., 2018).

The short version

Report AI-assisted thematic analysis the way you would report any other analytic tool: name it, cite it, state which phase it touched, and describe exactly how a human reviewed its output. Put this in your Method section in plain, specific language, not in a footnote or a limitations paragraph. Reviewers are not looking for a reason to reject AI-assisted work outright. They are looking for evidence that a human researcher, not the model, remained responsible for the interpretation.

Streamline Your Coding Phase: If you're using thematicanalysis.ai/analyze for the coding stage of your project, the platform simplifies this entire reporting process. You can save time and stress of thematic analysis by generating a first-pass codebook in minutes. Most critically for your methods section, the tool keeps every AI-suggested code linked to its source extract and tracks what you changed, which is exactly the kind of audit trail required by JARS-Qual. Export that log alongside your manuscript and the paragraph above writes itself from your own data rather than from a template.

References