Occurrence reporting systems in aviation ground handling hold records that have already been investigated, answered with a corrective action and closed. This preliminary single case study reports a retrospective comparison in which a large language model read all closed records of one ground handling station at Antalya Airport, Türkiye, over five years (N = 370), under a fixed prompt that supplied the recorded root cause and actions and told the model not to assume conditions absent from the record. Proposals were coded against the recorded analysis on a three-level convergence scale. Full or partial convergence was 76.5% for root causes and 65.9% for actions, and complete divergence was 10.5 percentage points higher for actions (paired McNemar test, p ≤ 0.009). A rerun on a stratified subsample of 74 records, coded independently by two coders (weighted kappa 0.92 for causes and 0.86 for actions), reproduced this profile only in part, converging more often with no difference between the dimensions (Stuart–Maxwell p = 0.30); none of its 545 output units was coded as introducing an absent condition. Seven experts judged the approach useful as a complementary decision support layer. The figures describe agreement between readings, not accuracy or improved safety.
Copyrights © 2027