I keep coming back to the idea that empathy and intelligence are the same muscle. Empathy is really theory of mind — the ability to model what's going on inside someone else's head — and that modeling is just intelligence pointed at other minds. The more of it a person extends, to other people and to other living things, the better the society they tend to build. It's the engine under the golden rule, or Michael Shermer's gentler version: the ask-first rule — find out how someone actually wants to be treated instead of assuming.
Sociopaths complicate the picture. Can you be fully empathetic and still purely selfish? I don't think so. A lack of empathy tends to be one of two things: a genuine defect in theory of mind, or plain laziness. It is so much easier to slap a label on someone, to assign blame and fault, than to do the harder work of standing in their shoes.
This is the part that nags at me about AI. If we eventually build minds modeled closely on human ones, do they inherit empathy by default, or not? There really seem to be only two attractor states for a superintelligence: empathetic and benevolent, or sociopathic and dangerous. Which makes understanding what actually makes a sociopath tick not just a psychology problem but maybe the most important safety problem we have. You can't reliably avoid building a sociopathic machine if you don't understand the thing you're trying to avoid.