Many of the proposed machine learning (ML) based network intrusion detection systems (NIDSs) achieve near perfect detection performance when evaluated on synthetic benchmark datasets. Though, there is no record of if and how these results generalise to other network scenarios, in particular to real-world networks. In this paper, we investigate the generalisability property of ML-based NIDSs by extensively evaluating seven supervised and unsupervised learning models on four recently published benchmark NIDS datasets. Our investigation indicates that none of the considered models is able to generalise over all studied datasets. Interestingly, our results also indicate that the generalisability has a high degree of asymmetry, i.e., swapping the source and target domains can significantly change the classification performance. Our investigation also indicates that overall, unsupervised learning methods generalise better than supervised learning models in our considered scenarios. Using SHAP values to explain these results indicates that the lack of generalisability is mainly due to the presence of strong correspondence between the values of one or more features and Attack/Benign classes in one dataset-model combination and its absence in other datasets that have different feature distributions.