Centering Knowledge Along the Responsible LLM Supply Chain: An Empirical Study & Multi-Stakeholder Taxonomy

Abstract

Existing CSCW and organization management literature suggests that knowledge is a central construct when developing, deploying, and using technological systems that are embedded in multi-stakeholder supply chains. Yet, it has rarely been the focus of comprehensive empirical investigations across LLM and broader AI supply chains. In this work, we draw on semi-structured interviews with 71 LLM practitioners to examine the knowledge required to produce responsible LLM systems, and how practitioners acquire it within LLM supply chains. Our findings not only reveal knowledge blindspots, but also a knowledge access and translation gap, where practitioners recognize knowledge needs but cannot fulfill them. We explain these gaps by showing how knowledge exchanges occur (proactively or serendipitously), which organizational arrangements and tools facilitate them (e.g., knowledge intermediaries), and which barriers undermine them (including limited visibility, lack of mutual understanding, and unclear responsibility for sharing and maintaining knowledge). We further synthesize knowledge needs into a multi-dimensional taxonomy that characterizes knowledge based on its abstraction, theme, and lens. We then discuss how this taxonomy can serve both as a practical resource for organizations of the responsible LLM supply chain to improve actionability and accountability (for example, by supporting precise communication, prompting self-reflection, and facilitating mutual blindspot identification), and as a methodological tool to reframe, revise, and disambiguate research and policy works on responsible AI notions related to knowledge. We hope to inspire future work in the CSCW community by outlining research opportunities to support practitioners in accessing, exploiting, and sharing relevant LLM knowledge across the supply chain.