I think in the next several years AI will start showing up in different specific applications. I don't think there are systems that allow cross-domain learning yet, but when they figure that out, that is when I would start worrying. A cross-domain learner could learn to be a chess whiz and could learn how to drive a car using the same knowledge base and apply that knowledge to completely novel situations. As of writing this, AI can't really to this (but we're pretty good at it).
As for sentient AI, I'm a weak AI person. I don't think machines will ever develop intentional behavior (consciousness remains a mystery to us). A good argument to illustrate this is John Searle's Chinese Room thought experiment. So, there will be no AIs with existential crises running around.
That being said, AI's ability to automate complex behaviors and the corruptibility of humans make AI dangerous. There is also the whole syntax versus semantic issue to deal with too. That's bound to create some potential issues when the machine understands something different than their programmer understands something. While its technically a "bug", such unintentional behavior could lead to devastating results if the machine is given a high degree of responsibility.