In order to protect humans from this power, the reinforcement values of intelligent machines must be for the happiness of all humans rather than narrow interests like corporate profits or military power.

This is described more fully at:

http://www.ssec.wisc.edu/~billh/g/super.html