Abstract for: Testing Human and AI Strategies of Concept-Link Generation for Causal Loop Diagrams. The Case of Biodiversity Programming Texts
CLDs are often the first approximations in the pursuit of more comprehensive models, and our understanding of systems relies on valid procedures for concept generation and polarized link identification. The paper focuses on factors influencing semantic and structural characteristics of CLDs, by comparing strategies based on AI prompting and human cognitive mapping techniques. The study aims to answer several introductory research questions regarding the semantic and structural characteristics of CLDs, as well as the quality of outputs. This includes questions regarding 'concepts' and 'links', their number and complexity. The studied models included are: ‘general,’ e.g. Large Language Models (LLMs) and ‘specialized’, e.g. Graph Neural Networks (GNNs). For each type, at least two models are to be compared to human-generated output. Since a plethora of AI models and tools is available, seemingly useful in the proper development of CLDs, the very architecture of AI models (unknown in proprietary LLMs) and the fundamental problem of AI hallucination challenge the validity of reasoning included in some CLDs. This makes further research necessary on the outputs generated by different models under a differentiated set of circumstances.