Blackbox in the sense that it's very difficult to detect what the outputs will be given a set of inputs currently. They definitely can be understood at higher levels of abstraction but I think it'll be awhile before the "emergent" behaviors these models sometimes display is well understood.