Department of Defense outlines Skynet and Terminator development
geek.com
geek.com
http://www.wired.com/threatlevel/2012/03/ff_nsadatacenter/al...
So if a machine is killing humans somewhere, it's because some human pressed the button.
So I agree with you, I find the thought of humans killing other humans much more disturbing.
That this is military research makes this outcome vastly more certain, as the systems would be weaponized from the beginning.
However, there's hope. They mention "big data", which means they're unlikely to ever get truly general AI, let alone the self-improving kind - "big data" is what you do when your system is stupid and your data is cheap, and if that's your situation, you're never going to be build a system that's smart, because that requires extracting as much information from minimal input as possible. AGI is not an incremental improvement on machine learning methods, it's a whole different beast. As long as they're fundamentally misguided from the start, they won't be much of a threat.
China, on the other hand, does have people working on the "right" parts of this problem, and that's a lot more concerning...
Ask about the motives. It wouldn't be an Asimov machine for starters