Thematic Analysis and AI Flags: Why Braun & Clarke Triggers Detection

Because you described a profoundly human, subjective qualitative process using the most formulaic passive boilerplate in modern academia. When you write that "transcripts were read and re-read," "initial codes were generated," and "themes emerged from the data," you are not writing an idiosyncratic account of your research; you are reciting an established script that appears verbatim across hundreds of thousands of dissertations. AI detectors like Turnitin and iThenticate do not understand qualitative epistemologies. They evaluate statistical probability, n-gram distribution, and perplexity across sequential text segments. Because your methodology chapter follows an identical passive structure with almost zero lexical variance from standard textbook descriptions, the mathematical probability of every consecutive token is exceptionally high. The detector flags the entire section as machine-generated not because you used AI, but because human academic conventions have trained graduate students to write like statistical language models.

This statistical trap is especially painful for qualitative researchers because it directly contradicts the methodological guidance of Virginia Braun and Victoria Clarke. In their foundational 2006 paper and their subsequent 2019 and 2021 works, Braun and Clarke repeatedly condemned the "cookbook" approach where researchers pretend themes passively "emerge" without researcher agency. Yet institutional pressure to sound "objective" often forces students into rigid, depersonalized prose. The result is a chapter that triggers every classifier flag while simultaneously violating the core tenets of Reflexive Thematic Analysis (RTA).

Resolving an AI flag in qualitative methodology requires understanding two interlocking systems: the algorithmic mechanisms of text classifiers, and the epistemological integrity of your audit trail. You cannot solve this problem by swapping words with generic synonyms. You solve it by grounding every generic step in your specific study context, restoring the active researcher voice, and presenting tangible documentary evidence of your analytical journey.

Braun and Clarke's six-phase framework—familiarization, coding, generating initial themes, reviewing themes, defining and naming themes, and producing the report—is the most widely cited qualitative analytical procedure in the social sciences. Introduced in their 2006 paper Using thematic analysis in psychology, the framework was designed to provide a systematic yet flexible approach to qualitative data. However, pedagogical replication has calcified this flexible guide into a rigid academic template.

"Thematic analysis is often treated as a singular, rigid method, with researchers adopting a cookbook recipe approach that ignores the theoretical and subjective choices underpinning Reflexive TA." (Virginia Braun & Victoria Clarke, 2019, Reflecting on reflexive thematic analysis, p. 591)

Thousands of graduate students across psychology, nursing, sociology, and education submit dissertations containing virtually identical paragraphs for each phase. Consider the textbook formula commonly found in student submissions:

• Phase 1 (Familiarization): "In the first phase, all interview transcripts were read and re-read multiple times by the researcher to achieve deep familiarity with the data."
• Phase 2 (Generating Initial Codes): "In the second phase, initial codes were generated systematically across the entire dataset, identifying meaningful segments relevant to the research questions."
• Phase 3 (Generating Themes): "In the third phase, codes were collated into potential candidate themes, clustering related concepts into broader overarching patterns."
• Phase 4 (Reviewing Themes): "In the fourth phase, candidate themes were reviewed against the coded extracts and the entire dataset to ensure coherence and validity."
• Phase 5 (Defining and Naming Themes): "In the fifth phase, themes were refined, clearly defined, and named to encapsulate the conceptual essence of each domain."
• Phase 6 (Producing the Report): "In the final phase, the analytic narrative was compiled, integrating illustrative empirical quotations to address the core research inquiries."

When text follows this exact six-point cadence, it exhibits zero informational surprise. It tells the reader—and the automated detector—nothing about what was actually done. It does not state what software was utilized, what conceptual disagreements arose, or how participant perspectives challenged the researcher's preconceptions. It is pure methodological boilerplate, designed to check a box rather than document an empirical inquiry.

To understand why this boilerplate triggers AI detection tools, you must examine how classifiers evaluate text. Most modern AI writing detectors, including Turnitin's proprietary model, analyze text sequences through statistical language modeling metrics, primarily perplexity and burstiness. Perplexity measures how predictable a sequence of words is given the preceding context. If a model can predict the exact next word with 98% confidence, the perplexity score is extremely low.

Large language models are trained on vast web corpora containing thousands of published academic papers, student dissertations, and online methodology repositories. In this training data, the transitional phrases of qualitative methodology are overwhelmingly uniform. After the tokens `"transcripts were read and"`, the probability distribution peaks overwhelmingly on `"re-read"`. After `"initial codes were"`, the top predicted tokens are `"generated"`, `"developed"`, or `"assigned"`. In the eyes of a statistical classifier, this predictable syntax mirrors the low-temperature generation of an autoregressive language model.

The problem is compounded by burstiness, which measures variance in sentence length, rhythmic structure, and syntactic complexity. Human writers operating in creative or argumentative registers naturally vary their pacing: short, declarative observations are followed by complex compound clauses containing qualifications and unexpected metaphors. But academic methodology writing traditions systematically suppress burstiness. Research supervisors urge students to maintain an unvarying, detached, objective cadence. When uniform sentence lengths are combined with predictable n-grams across two thousand contiguous words, the segment-level classifier outputs a solid band of red.

Turnitin's own engineering documentation confirms this segment-based behavior. As documented in their technical guidance, submissions are broken into overlapping text segments, each segment is scored for machine probability, and qualifying prose sentences inherit the pooled scores of overlapping windows. When an entire methodology chapter maintains one monotonous register, adjacent segments give the model nothing to differentiate. The resulting report highlights the section not because it caught a single telltale phrase, but because the entire linguistic texture matches the detector's model of automated text.

To break the low-perplexity profile without sacrificing academic rigor, you must dismantle the generic boilerplate and ground each of Braun and Clarke's six phases in your unique empirical reality. Reflexive Thematic Analysis requires transparency about the researcher's subjective analytical decisions. By documenting the concrete mechanics of your project—participant IDs, software nodes, coding anomalies, and conceptual pivots—you naturally generate high-burstiness, human-specific prose that no generic language model would predict.

Consider the direct comparison below, contrasting the standard formulaic description of each phase with a grounded, audit-ready version drawn from an empirical study:

Notice the dramatic difference in lexical density, burstiness, and specificity. The grounded versions employ first-person active voice ("I listened", "I generated", "I printed"), anchor the timeline to tangible artifacts (14 recordings, 142 open codes, NVivo 14, index cards), and explain the intellectual justification for theoretical adjustments. This prose cannot be mistaken for a template because it contains verifiable empirical facts unique to your dissertation.

If your methodology chapter has already been flagged by Turnitin or your supervisory committee, arguing about detector false-positive percentages alone is unlikely to persuade an academic integrity panel. Committees are not convinced by theoretical debates about algorithmic bias. They are convinced by an incontrovertible audit trail—documentary artifacts produced during the research lifecycle that no generative language model could fabricate retrospectively.

• The Codebook Evolution Log: Export your coding structure at multiple timepoints (e.g., initial open coding in Month 2 vs. consolidated axial themes in Month 5). If using CAQDAS tools like NVivo, MAXQDA, or Atlas.ti, export timestamped XML project histories, node hierarchy screenshots, and codebook change logs showing when nodes were merged, renamed, or retired.
• Timestamped Transcripts with Marginalia: Retain raw audio files with original creation metadata, matched against initial automated or manual transcripts. Maintain working drafts marked up with handwritten marginal comments, color-coded highlights, and immediate analytic impressions recorded during early analytical passes.
• The Reflexivity Journal (Reflexivity Memos): A comprehensive research diary documenting your subjective positionality, emotional reactions, role conflicts, and interpretive dilemmas. Dated entries recording why you rejected an advisor's thematic categorization or how your personal background influenced the interpretation of an ambiguous participant quote provide undeniable proof of human authorship.
• Supervisory Supervision Trails: Compile email correspondence, supervisor meeting minutes, and tracked-change Word drafts where specific coding frameworks, code definitions, and thematic structures were debated, reviewed, and progressively revised over months.

When preparing for an academic integrity review, assemble these four components into a chronological dossier. A committee examining a multi-month progression of timestamped NVivo exports, handwritten interview annotations, and personal reflexive memos can clearly see the cognitive labor of research. Against that documentary density, a 45% Turnitin indicator collapses from an accusation of academic misconduct into what it truly is: a statistical artifact produced by formulaic methodology prose.

When a thesis methodology section returns a high AI percentage, the instinctive reaction is often catastrophic panic. Students frequently make the critical mistake of feeding their entire chapter into generic web-based paraphrasers. This approach is fatal for qualitative dissertations: generic spinners destroy specialized methodological vocabulary, introduce distorted synonyms for established theoretical frameworks, and mangle verbatim participant quotations that cannot be altered under research ethics approvals.

A rigorous revision begins with surgical isolation. You must separate the text that requires contextual restructuring from the empirical foundations that must remain intact. Verbatim interview extracts, participant demographic tables, formal ethics approval numbers, and verbatim institutional citations should be strictly protected. Only the connective procedural prose—the formulaic transition paragraphs that carry low perplexity—should be rewritten.

For manuscripts with flagged sections, tools designed specifically for academic document integrity provide a structured workflow. Import the Turnitin or iThenticate report and rewrite only the passages it flags, returned as the same editable file. This targeted approach ensures that your underlying document architecture, reference citations, qualitative data tables, and participant interview transcripts remain completely untouched, while the flagged boilerplate sentences are recast into precise, human-authored academic prose.

Once the flagged segments are identified, execute revisions by replacing generic definitions with study-specific operational accounts. Replace passive summaries of textbook phases with active descriptions of your analytical choices. Clarify why you chose reflexive thematic analysis over content analysis or grounded theory, articulate your epistemological stance (e.g., critical realist or social constructionist), and describe the precise point at which data saturation or thematic sufficiency was achieved. This turns a vulnerable, generic methods section into an unassailable empirical narrative.

Why does Turnitin flag my verbatim participant quotes in qualitative findings? Turnitin's AI writing detector evaluates prose sentences across qualifying text and does not provide an automated toggle to exclude quotation marks, unlike its similarity index. When an interviewee speaks in complete, formal sentences or when an excerpt is translated into standardized English prose, the classifier treats those lines as standard qualifying prose and evaluates their statistical probability alongside the rest of the text. Keep raw interview audio recordings and timestamped transcripts ready to demonstrate that flagged quotes represent empirical field data.

Should I cite Braun & Clarke (2006) or their 2019/2021 works? Both, but with clear conceptual distinction. While their 2006 paper introduced the six phases, Braun and Clarke's 2019 and 2021 publications explicitly distinguish Reflexive Thematic Analysis (RTA) from coding reliability and codebook approaches. Citing their recent work demonstrates an updated understanding of researcher subjectivity, which provides theoretical justification for using active voice and personal reflexivity in your methodology chapter rather than generic passive boilerplate.

My supervisor insisted that I use passive voice in Chapter Three. What should I do? Explain that modern qualitative reporting standards, including APA 7th edition and Braun & Clarke's own reflexive framework, explicitly endorse first-person active voice ("I coded", "we constructed") to reflect researcher accountability. If your supervisory committee strictly enforces third-person passive voice, maintain the passive syntax while injecting granular empirical details—such as specific participant counts, software node figures, and concrete coding dilemmas—to disrupt algorithmic predictability without violating institutional formatting preferences.

Does using qualitative analysis software like NVivo or MAXQDA trigger AI detection? No. AI detectors analyze linguistic text patterns, not software metadata. Using NVivo, MAXQDA, or Atlas.ti does not trigger AI detection; rather, it produces the exact audit trail needed to defend against false positives. Exporting your project history, code hierarchies, and coding strip visualizations provides unassailable timestamped proof that your qualitative findings were developed through systematic human analysis.

Can I bring Turnitin's published false-positive rate into an academic integrity meeting? While Turnitin acknowledges a baseline false-positive rate under 1% for documents with over 20% AI writing and higher rates in specialized bands, citing general vendor limitations is rarely persuasive on its own. Panels view technical debates as deflection. Present vendor limitations only as context, while centering your primary defense on your physical audit trail: version histories, raw transcripts, coding evolution charts, and reflexivity journals.

If I rewrite my flagged methodology section, will my overall AI score drop to zero? No reliable service or tool can guarantee a specific numerical score on Turnitin, as detector algorithms and classification thresholds are proprietary and continuously updated. Rewriting flagged methodology paragraphs replaces low-perplexity boilerplate with authentic, study-specific narrative prose. This substantially reduces formulaic predictability and restores academic integrity, but aiming for an arbitrary "zero percent" metric misunderstands the statistical nature of machine learning classifiers.