Brain-machine interface (BMI) technology has rapidly matured over the last two decades, mainly thanks to the introduction of artificial intelligence methods, in particular machine learning algorithms. Yet, the need for subjects to learn to modulate their brain activity is a key component of successful BMI control. Blending machine and subject learning, or mutual learning, is widely acknowledged in the BMI field. Nevertheless, we posit that current research trends are heavily biased towards the machine learning side of BMI training. Here, we take a critical view of the relevant literature and our own previous work, in order to identify key issues for more effective mutual learning schemes in translational BMIs, specifically tailored to promote subject learning. We identify the main caveats in the literature on subject learning in BMI, in particular lack of longitudinal studies involving end-users and shortcomings in quantifying subject learning, and pinpoint critical improvements for future experimental designs.