The term "friendliness", when used of an AI, isn't intended to indicate friendliness in the same sense we mean it of people. Instead, it means that the AI has a deep and unchangeable disposition to treat humans well, regardless of how the AI evolves over subsequent generations.

An AI that is ignorant or uncaring of our interests - or one that could evolve in that direction - could do humans enormous damage.

The mission of the friendly AI project is to find ways to design AIs so that, even though AIs will be able to improve themselves and re-write their own source code, the AIs will maintain their "friendly" attitude at all times.