There will be a temptation to put AI in charge of launching nuclear weapons, to try to distinguish a real attack from a false alarm. That would be very dangerous.

In every false alarm so far, some human decided, "Never mind what this looks like — let's not destroy humanity on the supposition it is a real attack. How would we feel on finding out it wasn't one?"

An AI system would not reason this way, if it weren't programmed to.