Optimal and Private Learning from Human Response Data