Nosing Around: Dowsing the Data Landscape with an AI Divining Rod AI as a Search Instrument in the Preliminary Phase of Qualitative Analysis

Abstract

This article introduces the "AI divining rod," a procedure that uses large language models (LLMs) exclusively in the preliminary phase of qualitative analysis. By analogy with the "nosing around" of the Chicago School, the researching subject explores the interview material by roaming through it. The AI first identifies concepts in the scholarly literature selected by the author. After being checked against the original sources, they are formulated as questions (prompts). In the second step, based on these prompts, it provides summaries of the interview material, suggestions for categories, and, above all, supporting text passages for further analysis by the researching subject. The actual formation of categories and theory remains exclusively with the researching subject. The procedure is open regarding both method and tool, but it requires retrieval-augmented generation (RAG) systems; for reasons of data protection, local models are to be preferred. Against the backdrop of grounded theory, the AI divining rod is situated within the field of positions on LLMs: rejection, human-in-the-loop, AI-in-the-loop, and co-constructive approaches. It shares with most of these positions the premise that knowledge is actively constructed. It draws the line between indicating and interpreting, however, earlier than all the others. This difference in agency is traced back to the contrast between pragmatism and hermeneutics. An outlook shows that the interpreting subject can also be a participatory collective that includes co-researchers from the community. The benefit lies not in automation but in a division of labor that preserves the interpretive sovereignty of the subject.