There is no mechanism in this paper (or in a standard VAE, or GAN) to encourage that a single human-understandable semantic quantity should be captured in a single dimension of the latent code. So, in general, it won't happen. "Blondeness" is spread out over all dimensions of the latent code. Therefore the method in this paper (summarized by matheist) is totally reasonable. There are other autoencoder-type generative models that try to concentrate each attribute in one dimension, usually by using the class labels as an additional input. But that is not the focus of this paper.