This paper deals with the issue of features construction and selection for signals acquired during non-invasive Brain-Computer Interface (BCI) experiments. The so-called Echo State Network (ESN) architecture, a reservoir computing approach proposed by H. Jaeger in 2001, is first adapted to the specific issue of EEG signals classification. In order to predict the performed task, a commonly used ESN architecture is combined with regularized logistic regression trained following aggressive subsampling principles. The resulting method is shown to significantly outperform classification rates obtained using raw EEG signals. Basic single ESNs are then integrated to take advantage of ensemble learning techniques and aggressive subsampling principles. The resulting new architecture, called Uncooperative Democratic Echo State Community (UDESC), constitutes one of the first attempt to provide an efficient subject-independent features construction algorithm. Based on the generative power of individual ESNs as well as the discriminative abilities of ensemble learning combined with aggressive subsampling, it is shown to advantageously integrate the knowledge acquired by each single ESN. The results shown along this paper make an extensive use of a real training dataset made available to the BCI community during BCI Competition 2008. This dataset consists of four subjects involved in a two-class motor-imagery BCI experiments.