This work analyzes human behavior recognition approaches using WiFi channel state information from the perhaps less usual point of view of training and calibration needs. With the help of selected literature examples, as well as with more detailed experimental insights on our own Doppler spectrum-based approach for physical motion/presence/cardinality detection, we first classify the diverse forms of training so far employed into three main categories (trained, trained-once, and training-free). We further discuss under which conditions it is possible to move toward lighter forms of calibration or even succeed in devising fully untrained model-based solutions. Our take home messages are mainly two. First, reduced training might not necessarily kill performance (although, of course, trade-offs will emerge). Second, reduced training must come along with a careful customization of the technical detection approach to the specificities of the behavior recognition application targeted, as it seems very hard to find a one-size-fits-all solution without relying on extensive training.