In this case, the machine is learning, for example, that women do dishes and men drink beer. This isn't based on empirical data and patterns. It comes from the data the algorithm is trained on, which in this case is "...more than 100,000 images of complex scenes drawn from the web, labeled with descriptions". Those descriptions inevitably reflect human stereotypes (which again aren't just patterns). "Both datasets contain many more images of men than women, and the objects and activities depicted with different genders show what the researchers call “significant” gender bias."
Should machines understand that men are more likely to be construction workers than women, I think so. But that doesn't mean that biased data is not a huge problem.
At the very least, we should strive to teach machines to understand the world as it is, not as we view it through flawed, biased eyes. Datasets generated by humans are going to be invariably flawed and far from reality, unless we take careful steps to ensure otherwise. Emphasizing the importance of these datasets, we should be particularly concerned when we are teaching machines to act based on our biases and find that (in the case of this article), the machine is actually learning to amplify our own biases.