Brain-Computer Interfaces (BCI) are communication systems that enable their users to send commands to computers through brain activity only, this activity being measured and processed by the BCI (usually using ElectroEncephaloGraphy – EEG). Making computer control possible without any physical activity, BCI have promised to revolutionize many application areas, notably assistive technologies for paralyzed users (e.g., for wheelchair control) and human-computer interaction (HCI). Despite this promising potential, BCI are still barely used outside laboratories, due to a poor reliability. For instance, current BCI based on only 2 imagined movements correctly recognize less than 80% of the users’ mental commands, on average, while between 10 to 30% of BCI users (depending on the BCI type) cannot control a BCI at all. Designing a reliable BCI requires to consider it as a co-adaptive system, with its users learning to produce distinct brain activity patterns that the machine learns to recognize using signal processing. Indeed, BCI control is a skill that the user has to learn. Most research efforts so far have been dedicated to signal processing or human-computer interaction techniques, i.e., on the computer side. Unfortunately BCI user training is as essential but 1) is only scarcely studied and 2) standard approaches are only based on heuristics, without satisfying human learning principles. Thus, currently poor BCI reliability is probably due to a large extent to highly suboptimal user training. Therefore, to obtain a much higher reliability for BCI we need a major rethinking of their fundamentals in algorithmics (signal processing, machine learning) and user training (feedback and training tasks). In particular, we propose to create a new generation of BCI that apply human learning principles to ensure the users can learn high quality control skills which will go much beyond those obtained with currently available systems, hence making Brain-Computer Interfaces reliable and trustable. To do so, we will first work on understanding and modeling BCI skill acquisition from a neurophysiological point of view. In other words, we first aim at identifying what are the EEG features defining good EEG patterns (that are successfully recognized by the BCI), and how they evolve with training. Then, we will propose new EEG signal processing tools to quantify such training-related EEG features in real-time. This will enable us to identify objectives to reach with BCI training and a way to quantify and guide the user’s progress during training. Afterwards, we will combine these new EEG features and BCI training models with recommendations and principles from human learning and education psychology to propose new and relevant feedback and training tasks to radically improve BCI training. In particular, we will propose adaptive and adapted training tasks and provide the users with an explanatory feedback (indicating what is good or bad about the EEG patterns performed) based on our new training-related EEG features. Finally we will extensively evaluate and validate our new BCI, first on healthy users, then on a few motor-impaired ones. Overall, we target a new BCI design leading to a fast acquisition of reliable BCI control skills. Such a reliable BCI could actually positively change HCI as BCI have promised but failed to do so far.