The more precise version of "friendliness" is value alignment - Stuart Russell describes the problem well here[1], if you're interested.

1: http://edge.org/conversation/the-myth-of-ai#26015