Can We Responsively Code Qualitative Data with an LLM Aide? A Human-In-The-Loop Approach

This article examines a human-in-the-loop method for integrating large language models (LLMs) into qualitative coding while preserving its interpretive nature. Using GPT-4o and GPT-5 to code 36,000 words of interview transcripts, the authors combined deductive and inductive, nonexclusive coding. They found LLMs accelerated coding and produced mostly accurate results, but erred through keyword-driven misclassification and granularity issues. Critical human review mitigated these flaws, suggesting LLMs work best as coding aides, not replacements.