Nguyen-Trung (2025) — ChatGPT in Thematic Analysis: GAITA and the ACTOR Framework

TL;DR: Introduces GAITA (Guided AI Thematic Analysis) — an adaptation of King et al.'s Template Analysis where the researcher leads reflexively through four stages while systematically guiding GPT-4. Also proposes ACTOR, a five-element framework for combining effective prompting techniques. Published in Quality & Quantity, GAITA is distinguished from other protocol papers by its grounding in an established qualitative method (Template Analysis) and its explicit transparency about AI limitations.

Problem

The gap GAITA addresses is specific: existing AI-TA protocols either adapt Braun & Clarke’s phases (Naeem et al., Goyanes et al.) or propose new frameworks without established methodological grounding (CAAI, AbductivAI). What is missing is a rigorous adaptation of a specific, well-validated qualitative method that preserves its theoretical structure while incorporating AI assistance.

Template Analysis (King et al. 2018) provides the base: a hierarchical approach to thematic analysis that emphasizes systematic organization of codes into a structured template, iterative refinement, and the use of the template as a living analytic document. It is more structured than reflexive TA but more flexible than predetermined deductive coding — well-suited to AI collaboration because it provides clear organizational principles without closing off interpretive possibilities.

The second problem the paper addresses is practical: how do you actually prompt ChatGPT for qualitative work in a systematic, teachable way? Existing guidance on prompting is either too abstract (“use clear, specific prompts”) or too context-dependent to generalize. ACTOR provides structure.

Approach

GAITA: four stages adapted from Template Analysis.

Stage 1: Data familiarization. The researcher reads data personally and develops initial impressions. AI then produces summaries that the researcher checks against their impressions — surfacing gaps and confirming or complicating initial readings. This is not delegating familiarization to AI but using AI to verify and extend the researcher’s initial engagement.

Stage 2: Preliminary coding. AI generates first-pass codes based on researcher-directed prompts. The researcher interrogates, revises, rejects, and supplements. The design ensures that AI’s coding output is treated as provisional — hypotheses about what the data contains, not descriptions of what it contains.

Stage 3: Template formation and finalization. Researcher and AI iteratively develop and refine the coding template — the hierarchical structure of codes and their relationships. This is where Template Analysis’s distinctive contribution (explicit template structure) is operationalized with AI assistance.

Stage 4: Theme development. Researcher drives thematic synthesis, with AI assisting in pattern identification across codes. The researcher is explicitly the “reflexive instrument and intellectual leader” throughout.

ACTOR: five elements for systematic prompting.

• Anchoring — situate the model in the research context and dataset before any coding tasks
• Chaining — use sequential prompts that build on each other rather than isolated single prompts
• Tasking — specify exactly what the model should produce (codes, summaries, comparisons)
• Organizing — structure the output format (tables, lists, hierarchies) to facilitate review
• Reflecting — ask the model to explain its reasoning, making the basis of outputs inspectable

ACTOR is designed to make prompt engineering systematic and teachable for researchers without technical backgrounds — democratizing access to effective AI-assisted qualitative work.

AI’s Role

AI is positioned as a guided research assistant — systematically directed by the researcher at every stage rather than left to operate autonomously. The GAITA framework’s core design principle is that the researcher’s reflexivity and methodological authority are never suspended. AI contributes speed and pattern detection; the researcher contributes contextual understanding, theoretical grounding, and interpretive judgment.

This is a more structured version of the “researcher-directed AI” model that characterizes most of the corpus, distinguished by the Template Analysis grounding and the ACTOR framework’s systematization of prompting.

Epistemological Stance

Post-positivist with interpretivist elements. Template Analysis, which GAITA adapts, sits between small-q and Big-Q traditions — it is more structured than reflexive TA but attends to interpretive depth rather than just pattern counting. GAITA preserves this middle position.

The paper engages more seriously with trustworthiness and validity than most protocol papers, framing ACTOR’s “Reflecting” element as a transparency mechanism: if the researcher can see why the AI produced a particular code, they can evaluate its basis. This is a partial audit trail, not full reproducibility.

Rigor and Trustworthiness

The grounding in Template Analysis is the paper’s strongest methodological asset. Unlike papers that propose novel frameworks, GAITA inherits the methodological validation that Template Analysis has accumulated over decades of application. The adaptation preserves the method’s core logic while incorporating AI assistance.

The paper’s transparency about limitations is unusual and valuable (see below). Acknowledging structural problems with ChatGPT’s context window, inconsistency, and training data gaps makes the framework more credible, not less.

Limitations

The paper names its limitations explicitly, which is a methodological strength:

• Restricted context window makes processing large datasets in a single session difficult — researchers must chunk data across sessions, risking loss of cross-session coherence
• Inconsistent outputs require multiple prompt attempts for the same task — adding time that partially offsets efficiency gains
• Session limits force movement across workspaces, disrupting the iterative workflow GAITA requires
• Limited relevant training data for qualitative research means ChatGPT’s default understanding of “thematic analysis” may not match the researcher’s methodological commitments

These are structural limitations of current LLMs, not limitations of the GAITA framework per se. They will change as models improve, but they affect any framework that relies on ChatGPT’s current capabilities.

Connections

• llm-qualitative-research — broader landscape
• prompt-engineering — ACTOR systematizes and extends the prompting literature; the five-element framework is teachable
• computational-grounded-theory — Template Analysis sits in the same coding-and-categorization tradition; GAITA adapts it for the LLM era
• goyanes-chatgpt-protocol-2025 — parallel six-step protocol without Template Analysis grounding; compare the frameworks
• naeem-chatgpt-ta-steps-2025 — parallel Braun & Clarke-aligned guide; compare the two approaches to covering all six TA phases
• friese-caai-framework-2026 — CAAI makes the more radical argument that coding itself should be replaced; GAITA works within the coding tradition
• validity-trustworthiness — ACTOR’s “Reflecting” element is Nguyen-Trung’s contribution to the audit trail literature
• nguyen-trung-nita-2026 — Nguyen-Trung’s own non-coding successor to GAITA; uses the same Down East dataset; finds 5/6 themes converge; demonstrates how the same researcher moved from coding to narrative-based approaches