Abstract This paper advances a theoretical contribution, a triadic framework of human-AI-society interaction, and introduces a methodological innovation to support it. We argue that AI development is not a unidirectional pipeline from design to use but a recursive system of feedback loops across three interdependent spheres: practitioner-AI (design decisions), user-AI (application practices), and society-AI (evaluation and governance). To develop this theory, we propose an LLM-enhanced computational grounded theory framework that positions LLMs as analytic instruments within a human-led workflow. The framework comprises five phases: data construction, LLM-assisted topic generation, human-LLM topic refinement, computational confirmation, and theory generation. We operationalize the framework through a case study of public discourse about ChatGPT on Twitter (now X), analyzing 584,160 English-language tweets from the first three months following the model’s release. Empirical analysis reveals 10 themes and 54 subtopics that converge into three stakeholder-oriented interaction patterns corresponding to the triadic framework. The theoretical contribution is threefold: (1) reconceptualizing human-AI interaction as irreducibly triadic rather than dyadic; (2) specifying the mechanisms of recursive feedback loops (design enables use, use generates evaluation, evaluation feeds back into design); and (3) identifying temporal asymmetries (user adaptation in hours, corporate response in days, governance in years) as a structural feature of AI sociotechnical systems. Methodologically, this study demonstrates how LLMs can support saturation assessment at scale while maintaining interpretive rigor.
He et al. (Mon,) studied this question.