Of course we have an idea. Depends on the AI models as well. Assuming neural networks, it's not much difficult (and under certain circumstances - can be rigorously proven) that neural networks can "simulate" e.g. Kalman Filter, which again under certain circumstances is equivalent to LSE (least-square estimators). So unless you be more specific, it can be said that neural networks learn probability distributions in a general sense.

For instance, if I remember correctly, the result of “equating” KF and neural networks has been established in 1990s already.