Can large language models mimic airline passenger preferences?

Abstract

Large Language Models (LLMs) are being used in the air travel sector to simulate passenger behaviour. While commercial LLMs provide ready-made solutions, concerns over reliability limit widespread use. This paper leverages decades of discrete-choice research to audit 23 LLMs’ responses to a flight-ticket choice experiment with zero-shot prompting. The results of our logit regressions indicate that LLMs can simulate highly rational preferences, correctly sign ticket attributes, differentiate between relevant and irrelevant factors, and deliver plausible willingness-to-pay estimates. However, LLMs might fall short in replicating nuanced demographic segmentation and show sensitivity to cultural bias in their training data. Their outputs are best used to inform early-stage modelling, with traditional market research remaining essential for validation.