Аннотация:Active brain-computer interfaces (BCIs) offer agents possibilities of affecting the world in a way that is distinct from bodily actions. One of the major disadvantages of current non-invasive BCIs is slow response time to the user's command, due to the need to classify temporally extended brain processes. In this paper, I propose a hypothesis that if BCIs will gradually approach the speed and precision of voluntary movements, a variant of the “uncanny valley” effect may emerge. In current interface architectures, a BCI user has to perform an intermediate mental action to give a command to the computer. Several reported cases indicate that these actions, e.g. kinesthetic motor imagery, can be automated over time. I examine the structure of the agent's intentions before and after this automation and compare it to movement automation. I argue that the evolution of intentions in motor learning apply to BCI control learning, meaning that BCI control gravitates towards motor control. I consider the strict requirements for movements and their outcomes to produce the sense of agency with respect to bodily actions. These requirements include high congruence between intention and the outcome of action and small delays before the outcome. Based on the putative proximity of motor control and control over fast BCIs of the future, I claim that without an extremely high classification accuracy, actions performed with these BCIs will severely disrupt the sense of agency of the user.