Someone else, mentioned this, as well, but the evidence seems to be mixed, in any case. I responded elsewhere with an example of an analysis of real-world data showing a bimodal left-right distribution of real-world voters in U.S. states, implying that a single Gaussian certainly isn't the best model.