Human Reliability Analysis (HRA) evaluates the contribution of human errors to system risk. Organizational factors (OFs) play a critical role in shaping human performance and contributing to human error. However, many HRA methods represent OFs either as lists of influencing factors without explicit causal pathways or through causal structures that rely heavily on expert judgment. Although the effects of OFs on human performance have been extensively investigated across safety-critical domains, such as healthcare, energy, aviation and aerospace, and maritime and offshore operations, the relevant evidence is dispersed across accident investigations, regulatory reports, and research publications. This evidence is predominantly unstructured, often expresses causality implicitly and contextually, and may describe similar relationships using different terminology, making systematic identification and synthesis burdensome. To address this gap, our broader research aims to develop a generic, data-driven causal model of organizational influences on human performance. As part of the broader research effort, this paper presents a large language model (LLM)-based pipeline for extracting textual evidence of causal relationships among OFs. A corpus of 266 publications collected in a previous study was processed using PDF-based text extraction and a structured seven-step procedure specifying the extraction task, output schema, formatting rules, causal terminology, sample outputs, definitions of 19 OFs, and 38 target causal relationships. The OF definitions were developed through expert-led literature review and knowledge-based synthesis, while the target relationships were identified a priori based on expert knowledge.
Three LLMs were evaluated using a common extraction task. GPT-5-mini was selected for full-corpus processing because it achieved extraction quality comparable to GPT-5 and Grok-4 at lower computational cost. The pipeline identified 1,538 causal instances, of which 1,482 (96.4%) matched the predefined target relationships. Evidence was found for all 38 relationships with frequencies varying substantially, such as seven occurred more than 50 times, 24 occurred 20–50 times, and seven occurred fewer than 20 times. Information exchange and coordination affecting Decision framing and execution and Employee training and development affecting Knowledge and abilities were the most frequently identified relationships, with 116 and 106 instances, respectively. The extracted causal evidence and frequency-based importance rankings were used to support construction of a Bayesian belief network causal structure. Through further prompt engineering and literature supplementation, the pipeline was also extended to extract evidence describing how different factor states contribute to human-performance degradation, supporting probabilistic model development presented in a companion study. Limitations included inconsistent factor naming and occasional weak or incoherent supporting excerpts. Overall, the results demonstrate that structured LLM extraction can transform dispersed qualitative evidence into a traceable dataset for reviewing OF causal models and supporting Bayesian network development.