Data assimilation allows for updating state variables in a model to represent the initial condition of a catchment more accurately than the initial Open Loop simulation. In hydrology, data assimilation is often a pre‐requisite for forecasting. According to Hornik [1991], artificial neural networks can learn any nonlinear relationship between inputs and outputs. Here, we hypothesize that neural networks could learn the relationship between the simulated streamflow (from a hydrological model) and the corresponding state variables. Once learned, this relationship can be used to obtain corrected state variables by applying it to observed rather than simulated streamflow. Based on this, we propose a novel, ensemble‐based, data assimilation approach. As a proof of concept and to verify the above mentioned hypothesis, we used an international testbed comprising four hydrologically dissimilar catchments. We applied the new data assimilation method to the lumped hydrological model GR4J, which has two state variables. Within this framework, we compared two types of neural networks, namely Extreme Learning Machine and the Multilayer Perceptron. Using well‐known metrics such as the Continuous Ranked Probability Score, we compared the assimilated streamflow series with the Open Loop streamflow series and with the observed streamflow. We show that neural networks can be successfully used for data assimilation, with a noticeable improvement over the Open Loop simulation for all catchments.