This article deals with a method for seismic damage identification in buildings with steel moment-frame structure. The damage identification is based on artificial neural networks and natural frequencies. A simplified finite element model is used to obtain the data needed for training the nets. The method is simulated on a four-storey building under conditions as close as possible to reality. The robustness of the method and its sensitivity to the variations of the mass with time and the influence of the data errors is addressed. The statistical analysis of the results is successful, but it reveals that the predictions are quite sensitive to the data errors.