Here's a way of connecting the views of "social morality" and "group morality", and explaining why groups using expected value wouldn't result in society using expected value. (Not that I think it should, but that's a different discussion.)

Say society is composed of different groups, each with their own coordination problems, and behaviors that could solve them. Say you can analyze a situation x in the space X, and for each group g, find the gradient ∇u(g,x) of their utility surface u(g,X). Each group g then prefers an action given by the direction of ∇u(g,x), with a strength of preference given by its magnitude.

Suppose each agent z is a member of one group g(z). (We don't need this assumption, but it makes the notation simpler.) If we call "socially-moral behavior at x" the average, over all agents z, of ∇u(g(x),x), I'm pretty sure this is going to give results that do not maximize social expected value.