Any learning algorithm works by finding some patterns in the data. In unsupervised learning, this usually means finding groups of instances which share similar values for some of their features. So potentially the
unknown value could be used by the algorithm as part of a pattern, and this could be a problem because it doesn't represent a real semantic information about the data.
This is what the author means, but this problem is unlikely to happen with a decent dataset: if the dataset is large enough, then the other features are likely to be different (have a lot of diverse values) in the subset of instances which have
unknown for feature X. In this case the algorithm is unlikely to consider them as sharing a pattern. If it does rely on this weak pattern, then it would mean that there is no other stronger pattern in the data to rely on, so the task is unlikely to be very successful anyway.