Leveraging Large Language Models for Automated Export Control Screening: Evaluating LLMs Framework

Export control (EC) compliance is a critical yet labour-intensive process within research institutions, where the classification of sensitive technologies and cross-border disclosures often depends on expert interpretation of complex legal frameworks.This paper investigates the potential of large language models (LLMs), specifically in this study , to support EC screening through a multistage, expert-in-the-loop framework.The methodology includes prompt variation, regulatory conditioning, reflective reasoning, and expert-informed evaluation to simulate real-world compliance workflows.Using a curated dataset of UK research project descriptions and the UK Strategic Export Control List, we assess model performance across over 1,400 outputs.Results show that while both models benefit from domain-specific grounding, ChatGPT-4o consistently produces more stable and interpretable classifications.Prompt sensitivity, bias behaviour, and ambiguity handling are also examined to highlight model limitations.The findings suggest that LLMs can support early stage EC assessment but require structured prompting and human oversight to ensure regulatory alignment.