Aircraft approach flight path safety management provides procedures that guide the aircraft to intercept the final approach axis and runway slope before landing. In order to detect atypical behavior, this paper explores the use of data generative models to learn real approach flight path probability distributions and identify flights that do not follow these distributions. Through the use of Generative Adversarial Networks (GAN), a GAN is first trained to learn real flight paths, generating new flights from learned distributions. Experiments show that the new generated flights follow realistic patterns. Unlike trajectories generated by physical models, the proposed technique, only based on past flight data, is able to account for external factors such as Air Traffic Control (ATC) orders, pilot behavior or meteorological phenomena. Next, the trained GAN is used to identify abnormal trajectories and compare the results with a clustering technique combined with a functional principal component analysis. The results show that reported non compliant trajectories are relevant.