Contextual prompting and large language models in reflexive thematic analysis: what changes and what remains human

Reflexive thematic analysis (RTA) constructs patterned meaning through researchers’ situated engagement with data, making claims that large language models (LLMs) can perform it conceptually difficult. We examined whether contextual grounding changes model suggestions when methodological guidance is held constant. We compared our human analysis of twelve interviews with Vietnamese university English-language teachers with outputs from two model conditions. Both received identical phase-specific RTA instructions, and only the context-informed condition received a factual briefing on Vietnamese tertiary education. Across three runs per condition, contextual grounding increased the median proportion of labels linked to extracts from 81% to 86% and labels accompanied by rationales from 71% to 79%. It produced more setting-specific suggestions but also unsupported assumptions about hierarchy and face-saving. Neither condition reproduced the situated judgement used to adjudicate competing interpretations. We conclude that model outputs may inform critical reflection but cannot substitute for reflexive interpretation.