Modern machine learning methods have made enormous scientific progress in recent years and have fundamentally altered entire scientific disciplines. Particularly, artificial neural networks have played a crucial role in this process. This change is also apparent in hydrology, and numerous studies in recent years have shown that neural networks are among the best available models for a wide range of hydrological problems. However, in contrast to their excellent predictive performance, is the theoretical understanding of neural networks still insufficient, and questions about their optimal architecture or training remain up-to-date unanswered. This lack of a more fundamental understanding has also implications for hydrology and it is, for instance, still unclear whether neural networks are capable to extrapolate to unseen system states and to what extent they can be applied as learning tool. Given the vastly growing use of neural networks in hydrology, more research is needed that focuses not only on finding ever more use cases but also on a better understanding of uncertainties in network-based predictions to improve our knowledge about their limits and promises. The overall goal of the proposed research project DETOX (Demystifying recurrent neural networks in a hydrology context) is therefore to shed light on the black box of neural networks, test to what extent they are more than just excellent interpolators and investigate how we can use neural networks as learning tools.