Designing AI for Qualitative Research: Predictability, Transparency, and Control in AI tool Skimle.com

Abstract

Despite the rapidly growing capabilities of large language models (LLMs) and their adoption across diverse forms of knowledge work, their reception in the domain of qualitative research has been ambivalent. The same technologies that promise to let researchers work faster, draw on larger corpora, and surface patterns that would be impractical to find by hand also raise the prospect of analyses that are unreliable, opaque, and disconnected from the interpretive judgment. In this article, I first set out the central criticisms and risks that attach to the use of artificial intelligence in qualitative analysis, focusing on hallucination and the more insidious problem of cognitive offloading, whereby tools that automate interpretation gradually erode the researcher's own engagement with the data. Addressing the concerns, I introduce three design criteria for AI-driven analysis and their implementation in Skimle.com service: (1) predictable processes; (2) two-way, end-to-end transparency between original data and conclusions; and (3) substantive researcher control over AI-powered coding and categorization. I elaborate why these design characteristics enhance the ability of tools to augment rather than automate expert work. Finally, I consider the relationship between AI analysis tools and qualitative methods, arguing that a well-designed tool should remain method-agnostic, allowing researchers to follow their chosen approach.

Keywords: AI augmentation, qualitative research, large language models, CAQAS, cognitive offloading