Cognition and Computation in Decision-Making

Authors:

Abstract
This paper examining how local Large Language Models (LLMs) can partially automate the Critical Decision Method (CDM) in aviation safety investigations. The CDM, while widely respected for its ability to elucidate human factors and decision-making processes in rare or complex scenarios, often requires labor-intensive qualitative coding. To address this challenge, we developed a pipeline employing two specialised models: \emph{Phi-3-Mini-Instruct} for generating structured responses and \emph{Zephyr-7B-Beta} as a “judge” to evaluate confidence, completeness, and groundedness. A single anonymised incident served as our pilot case. Seventy-two participants (36 aviation professionals (pilots) and 36 novices) responded to a 53-item CDM-inspired questionnaire, creating a human reference dataset. The pipeline’s performance was benchmarked against both this human data and a classical NLP baseline (TF-IDF + SVM). Results revealed that the LLM matched 78\% of the majority-human multiple-choice answers and achieved a mean absolute error (MAE) of 0.38 on Likert-scale questions. Its open-ended responses, although moderately accurate, occasionally exhibited factual hallucinations (e.g.\ referencing non-existent systems) and role misattributions. Further stratification showed that the LLM outperformed novices but did not match pilots’ domain expertise, underscoring the importance of operational familiarity for nuanced decision analyses. Despite the single-incident scope limiting statistical generalisation, these findings suggest that LLM-based tools can substantially expedite repetitive data processing and facilitate consistent categorisation tasks that often consume investigators’ bandwidth. Future work will expand to multiple incidents, integrate flight data recorder (FDR) and cockpit voice recorder (CVR) information to reduce speculation, and refine both self-evaluation mechanisms and ethical safeguards.