Decomposing Complexity: Modeling NLP Systems for Crisis Sensemaking with Systemigrams

Abstract

Generative artificial intelligence (GenAI) tools, such as Large Language Models (LLMs), are increasingly deployed in high-stakes domains to support decision-making, situational awareness, and knowledge synthesis. Yet their integration into crisis workflows raises significant challenges related to interpretability, epistemic trust, and organizational alignment. Drawing on semi-structured interviews with intelligence and military professionals, this study investigates how practitioners evaluate and operationalize LLMs in environments defined by uncertainty and time pressure. Thematic analysis revealed three core tensions: Uncertainty and Mistrust of Black Box Systems in Crisis Situations, Interpretability and Operational Pressure, and Using AI to Support Sensemaking. These themes serve to motivate a systems-oriented approach to effectively integrating AI systems into sensemaking applications. Building on complexity theory and sensemaking literature, we propose a systemigram-based framework for modeling the interdependent components of NLP systems, Human, Prompt, LLM, Uncertainty, and Sensemaking, as a modular sociotechnical system. This framework provides a tool for both analyzing and designing NLP applications in crisis contexts, helping practitioners trace feedback loops, understand dependencies, and promote resilient, interpretable AI integration. Our findings underscore the importance of epistemic alignment and system-level transparency for effective human-AI collaboration in critical decision environments.

Ähnliche Arbeiten