DeTAILS: Deep Thematic Analysis with Iterative LLM Support

TL;DR

DeTAILS addresses the challenge of scaling thematic analysis by embedding LLM assistance into a six-phase, codebook-driven workflow grounded in Braun and Clarke’s reflexive thematic analysis. It couples a transparent, bidirectional human–AI loop with a boundary object (the codebook) and a local/hybrid architecture to preserve analytic agency while accelerating coding and theme development. Empirical results from 18 qualitative researchers show strong alignment between AI outputs and researcher refinements, along with substantial efficiency gains and positive usability signals, though researchers still rely on human interpretation to preserve depth and voice. The work advances design implications for trustworthy AI-assisted qualitative research, emphasizing opt-in AI involvement, uncertainty visualization, and deliberate scaffolding of reflexivity, with clear directions for future refinement and broader data modalities. It demonstrates that AI can extend qualitative inquiry by scale without eroding interpretive commitments when designed to foreground researcher judgment and transparency.

Abstract

Thematic analysis is widely used in qualitative research but can be difficult to scale because of its iterative, interpretive demands. We introduce DeTAILS, a toolkit that integrates large language model (LLM) assistance into a workflow inspired by Braun and Clarke's thematic analysis framework. DeTAILS supports researchers in generating and refining codes, reviewing clusters, and synthesizing themes through interactive feedback loops designed to preserve analytic agency. We evaluated the system with 18 qualitative researchers analyzing Reddit data. Quantitative results showed strong alignment between LLM-supported outputs and participants' refinements, alongside reduced workload and high perceived usefulness. Qualitatively, participants reported that DeTAILS accelerated analysis, prompted reflexive engagement with AI outputs, and fostered trust through transparency and control. We contribute: (1) an interactive human-LLM workflow for large-scale qualitative analysis, (2) empirical evidence of its feasibility and researcher experience, and (3) design implications for trustworthy AI-assisted qualitative research.

Figures (6)

• Figure 1: The DeTAILS workflow comprises six phases inspired by Braun and Clarke's Reflexive Thematic Analysis. Each phase requires collaboration between the researcher and LLM.
• Figure 2: Background Research: DeTAILS supports early-stage analysis by surfacing related concepts and building a structured concept outline, providing a grounded starting point for subsequent coding.
• Figure 3: Coding Phase: LLMs generate preliminary codes, which are then refined by researchers into an initial codebook to guide analysis across the corpus.
• Figure 4: Phases 4 and 5: Reviewing Codes and Generating Themes. These phases support higher-order interpretation, allowing researchers to refine code clusters and develop coherent, researcher-driven themes.
• Figure 5: Box and whisker plot of F1 Scores across each phase of reflexive thematic analysis. Background Research and Coding are measured by a weighted F1, whereas Reviewing Codes and Generating Themes use a Macro F1.