Who Does What? Archetypes of Roles Assigned to LLMs During Human-AI Decision-Making

TL;DR

This paper tackles the socio-technical challenge of assigning roles to large language models in human-in-the-loop decision-making. It develops a taxonomy of 17 human-LLM archetypes through a scoping literature review and thematic analysis of 113 studies, then evaluates their potential impact in a real-world clinical radiology case using a controlled prompt perturbation approach. The study finds that archetype selection can affect prediction accuracy, agreement with external references, and explanation quality, highlighting tradeoffs in decision control, knowledge sourcing, and cognitive workload. It offers design guidelines to support responsible deployment of LLM-supported decision-making systems and notes significant implications for safety, trust, and accountability in high-stakes domains.

Abstract

LLMs are increasingly supporting decision-making across high-stakes domains, requiring critical reflection on the socio-technical factors that shape how humans and LLMs are assigned roles and interact during human-in-the-loop decision-making. This paper introduces the concept of human-LLM archetypes -- defined as re-curring socio-technical interaction patterns that structure the roles of humans and LLMs in collaborative decision-making. We describe 17 human-LLM archetypes derived from a scoping literature review and thematic analysis of 113 LLM-supported decision-making papers. Then, we evaluate these diverse archetypes across real-world clinical diagnostic cases to examine the potential effects of adopting distinct human-LLM archetypes on LLM outputs and decision outcomes. Finally, we present relevant tradeoffs and design choices across human-LLM archetypes, including decision control, social hierarchies, cognitive forcing strategies, and information requirements. Through our analysis, we show that selection of human-LLM interaction archetype can influence LLM outputs and decisions, bringing important risks and considerations for the designers of human-AI decision-making systems

Figures (9)

• Figure 1: Scoping literature review methodology to generate human-LLM interaction archetypes, with an example paper coding provided. HITL = human-in-the-loop.
• Figure 2: 17 distinct archetypes were identified during scoping literature review of human-LLM decision-making, indicating the diverse socio-technical patterns used to engage humans and LLMs in shared reasoning and deliberation. Explainer, minority opinion, and implicit reasoner reflect subtypes of other archetypes.
• Figure 3: Archetypes can be combined, layered, and integrated within human-LLM decision-making contexts.
• Figure 4: a) The case study followed an overall methodology where archetypes were translated to task-specific prompts, appended independently to 100 different patient reports, then queried to an LLM, with evaluation metrics calculated and compared across each archetype. b) The translation of archetypes to prompts often required the addition of reference information sources (details about who and what external predictions were made) and variations in default conditions (whether the reference prediction was positive or negative). These details shaped which evaluation metrics were calculated and compared across each archetype.
• Figure 5: The selection of human-LLM archetypes involves assignment of varying autonomy levels to human and AI decision-makers. Solid lines demonstrate archetypes that involve direct subtypes, while dashed lines indicate archetypes that may be related or deployed simultaneously.