With the proliferation of new modeling techniques including Artificial Intelligence and Machine Learning algorithms and availability of large datasets to train models it is important to detect situations when model uses input outside of the training set. If the input is outside of the training \ set the results are less reliable. Historically the problem of extrapolation was considered for linear models and generalized linear models. In these cases the functional form of the model is well known and relatively simple ideas provide good guidance for extrapolation detection. The situation is different in the case of “black box” model with many more parameters or features that were typically used for the training of traditional models. This article proposes metric to detect extrapolation. This metric takes into consideration the original training set and the resulting model while treating the model as a “black box”. The metric is computationally simple and the model user can