Interesting thing about the default image, "beard" on a woman creates a more typically "masculine" jawline but doesn't add facial hair.
Increasing blondeness also increases smile.
Increasing blondeness also increases smile.
Just as an aside here: "Blondeness" and "beard" are probably just labels the authors found correspond the most to the latent variables in this case. This means that there won't be a perfect translation between those words and what these variables directly respond to in the network.
So although the training data may have been biased with more smiling blonde people, it doesn't necessarily have to have been so. It might be that what this latent variable encodes just does something else in edge cases where there are few examples.
See their code snippet halfway down their page, or "Semantic Manipulation" on page 8 of their linked paper.
Our negative so far is that most of the interpolations really seem more like two existing pictures were photoshopped together than like a new face was generated from a latent space and knowledge of faces. Sorry, I don't have the vocabulary and concepts of visual composition to say why. It just looks "shooped".