It is fascinating. Particularly the "characters" problem you mentioned already falls under the realm of natural language processing, and a very specific problem at that (context recognition).
I concede that defining "logical leap" is mostly a semantic argument -- I personally consider a syllogism (A->B, B->C therefore A->C) as a "logical leap", but it is one that can be perfectly captured by inference engines. Furthermore, one may argue that some aspects of "human intuition" might actually be... illogical.
And to answer your hard-coding concern: it ISN'T necessarily needed to hard-code more rules -- that is just a matter of making the system perform better given similar questions in the future. Like I said earlier, modern inference engines do have conflict-resolution mechanisms, as it is very possible in large knowledge bases that many rules are applicable to a particular case. One of the main research areas, I think, is for conflict-resolution mechanisms that can correctly recognize context as opposed to some "simplistic" or "hard-coded" statistical formula. (But what is the difference, really? One may argue that "context" is just more parameters to the formula.) Sure, humans are still quite ahead of A.I. in this regard, but the Mycin system is actually notable in that it was able to diagnose diseases MUCH better than humans can -- and that's in the 1970s.