I’ve been reading up on this area a bit and I haven’t yet seen a paper detailing the shortcomings or limitations of this approach to AI.  I’ll read more to see if someone has researched this…