Improving Decision-Making of Civilian and Military Pilots Under Pressure

An Artificial Intelligence-Based Approach Using Prospect Theory

Abstract

The expansion of Artificial Intelligence (AI), especially Large Language Models (LLMs), into safety-critical fields such as aviation raises fundamental questions about alignment with human expertise under risk. This study presents a controlled comparison of decision-making behaviour across professional pilots, lay participants and a LLM, each exposed to six hypothetical scenarios of Prospect Theory dilemmas and operational emergencies. A structured questionnaire measured risk preference, ethical reasoning, and context sensitivity; group differences were analysed using standard statistical methods. Results show that, in loss- and gain-framed dilemmas, both humans and the LLM display cognitive biases consistent with Prospect Theory. However, this mimicry fails to generalise: in operational risk scenarios, pilots consistently select the safest response, reflecting internalised safety culture and rapid pattern recognition. Lay participants show intermediate, more variable behaviour. By contrast, the LLM persistently selects risk-seeking, utilitarian strategies in operational domains, irrespective of prompt or training. The ethical dilemma scenario exposes further divergence: while nearly all humans act according to a deontological duty of care, prioritising those under their protection, the LLM favours maximising outcomes even when this increases risk for all. These findings show that current LLMs, despite surface-level alignment in abstract tasks, lack the tacit, context-dependent, and ethical knowledge needed for trustworthy deployment in safety-critical domains. Effective AI integration requires multi-layered alignment, expert feedback, explicit ethical constraints, and ongoing validation against expert practice. These results underscore the need for domain-specific alignment before adopting AI in aviation and related fields.