In this paper, the problem of concept drift detection in data stream mining algorithms is considered. The autoencoder is proposed to be applied as a drift detector. The autoencoders are neural networks that are learned how to reconstruct input data. As a side effect, they are able to learn compact nonlinear codes, which summarize the most important features of input data. We suspect that the properly learned autoencoder on one part of the data stream can then be used to monitor possible changes in the following stream parts. The changes are analyzed by monitoring variations of the autoencoder cost function. Two cost functions are applied in this paper: the cross-entropy and the reconstruction error. Preliminary experimental results show that the proposed autoencoder-based detector is able to handle different types of concept drift, e.g. the sudden or the gradual. © 2020, Springer Nature Switzerland AG.