Suppose 10 users on average interact with 2 non-ugly persons and 1 ugly person. People like commenting on the non-ugly people's content with "wow so pretty!" and "that's awesome! ", etc, etc while ugly people don't get as many comments and maybe even receive neutral to non-positive comments.

Now a new person signs up. They get recommend non-ugly people in their feed since that's more popular based on views and interactions.

Another new person signs up and they get the same recommendation, and so on.

After 100 new sign ups, the recommendation engine has 'learned' that majority of people prefer interacting with non-ugly people.

Another new user signs up and all they see in non-ugly people recommendations.

The end result is pretty much the same. Ugly people will get pushed out enough either by the programmatic learning engine that becomes over trained and biased, or by manual reviewers that filter content based on data that shows that non-ugly people bring in more users, otherwise they'd promote ugly people content if that was driving more interactions.