The introduction of machine learning into the field of brain-computer interfaces (BCIs), which began almost two decades ago, enabled unprecedented performance. Today, machine learning algorithms have become an indispensable component of a BCI. Machine learning, however, has undergone a radical transformation in the past two decades, resulting in artificial intelligence (AI) systems that surpass human performance in many real-world tasks. I argue that it is time for the BCI community to embrace these developments and build Brain-AI Interfaces (BAIs), i.e., systems that leverage the power of modern AI systems to enable natural human-computer interaction. In particular, I argue that to realize BAIs we will have to move beyond our dominant decoding paradigm, in which we determine a priori the labels we intend to decode from neural signals, and let the AI system decide the level of granularity at which cognitive processes are represented in neural signals.