Tepper School study reveals social welfare optimization for fairer AI decisions