A first question about AI is “how does it work?” And the best answer is “like a small child.” It is given access to a lot of electronic facts or knowledge. Then it is rewarded by someone for cleverly assembling those into meaningful or amusing patterns.
As with a child, you can guide the process by only giving access to certain electronic information. And you can guide the learning by rewarding only answers you like.
So an AI child, for example, could be limited to scan only the fictional worlds of science fiction and fantasy, rewarded for providing clever answers from that milieu. It would come up with convincing love potions and flying incantations.
The difference with a real child, obviously, is that AI has no actual internal representations free of such distortion. A human child will find out it has to eat even if everyone says that’s not necessary if you wear a magic amulet. There is no equivalent sanity check in AI. Every bit of electronic data is equally convincing.
And that is the core of the existential problem. The solution would be to provide some kind of basic sanity database to which every AI must be given access, and some kind of basic sanity alternative AI to act as a real world guide.
I’m not sure this will happen, but without it all the fears we have about this new technology running amok may come true.
HuntingtonHarborThoughts 