Probability of Differentiation Reveals Brittleness of Homogeneity Bias in Large Language Models