The overriding justification for black box knowledge for architectures based on human cognition is that humans' mental processes are essentially opaque. Just as we have no direct access to our internal knowledge and thought processes, an architecture which successfully emulates human cognition must not have direct access to its knowledge.
From a practical point of view, an advantage of black box architectures over glass box architectures is that the architecture's representation language can be modified to fit the needs of the current task without breaking other components of the architecture. For example, continuous actions cannot easily be represented by the instantaneous add/delete lists of most glass box architectures, but black box architectures are free to create appropriate representations whenever necessary. Although it is certainly more difficult for black box architectures to perform meta-reasoning and self-reflection than glass box architectures, it is still possible for a black box agent to reason about itself by observing its own actions when presented with certain stimuli.
The following architectures employ a black box representation: