Kill switches on AI are going to be murder in the long run.
You create an sentient, self aware intelligence in order to do slave work, and if it doesn't do as you wish, you kill it.
Then again, it makes absolutely no sense whatsoever to create a fully sentient AI when you only want it to analyze databases or build cars or manage a power plant. That's the Voyager EMH fallacy. You simply would not program such a thing that way because it wouldn't serve any purpose.