It depends on what we're referring to. There is no one all-encompassing AI, and that would be very difficult to implement even if we wanted to. But AI systems can break things, not out of a desire to do harm, but because they are not configured correctly and people come to expect them to be perfect. We're human, we learn as we go. Sometimes we have to learn the same lessons over and over.
But, yes- most of the time, the quick and easy solution is to turn it off. Very few AI systems will be able to prevent that, unless we program them to.