From delegation to moral abdication: classifying large language model uses by judgment and epistemic control

Abstract Ethical assessments of large language model (LLM) use are often organized around sectors, tasks, or abstract notions of societal risk. While valuable for governance and regulation, such classifications frequently fail to capture ethically salient differences between concrete practices of use. This paper develops a complementary, use-based taxonomy for analyzing LLM applications according to how judgment and epistemic authority are distributed between users and systems in concrete workflows and prompting practices. The taxonomy is structured along two dimensions: the degree of delegated judgment and the degree of epistemic control retained by users. It is operationalized through prompt-level analysis, treating prompts as delegation artifacts that encode how cognitive and normative labor is distributed between users and LLMs. Applying the taxonomy to two case studies—automated grading in education and contract termination in public administration—the paper shows how prompt-level analysis brings into view patterns of delegated judgment and epistemic control that remain obscured by sectoral, task-based, or interface-level descriptions, and how interface design and institutional framing can obscure responsibility-related vulnerabilities arising from extensive judgment delegation. Building on these analyses, the paper identifies the structural conditions under which human users can meaningfully exercise responsibility in human–LLM interaction. It develops an ethical framework for analyzing responsibility (understood in terms of knowledge and control) within concrete practices of LLM use. The central normative claim is that it is a necessary (though not sufficient) condition for ethically defensible LLM use that this use remains structured as tool use, preserving epistemic access, independent judgment, and justificatory authority. By complementing sector-, task-, and risk-based approaches with a focus on concrete practices and workflows, the taxonomy supports a differentiated ethics of LLM use and more precise public debate about LLM deployment across application domains.