A recent article, "Weight bias may harm obese children," summarizes a July 2007 Psychological Bulletin article:
When the study participants were asked to rank the children in the order of whom they would like to be friends with, they ranked the overweight child last. … Some studies found that a sizable number of teachers harbor negative views of overweight students, seeing them as "untidy," for example, or less likely to succeed than their thinner peers. Other research found that overweight children often report teasing from family members, including parents.
The article repeatedly uses the word "bias" to describe these negative judgments, but it doesn’t bother to show why these effects are in fact biases. Unless you want to claim that no one should ever be teased or prefer some as friends over others, or that teachers should never estimate student tidiness or success, the question is: what is the evidence that judgments made about fat kids are in fact too negative on average? Without such evidence, these negative judgments should not be called "biases."
"Spontaneous arbitrary group formation" could be a result of within-group competition for status. Attacking a group member with low status signals that you have more status than the member that was attacked. (The high status kids in school aren't the ones that pick on the outcasts. They don't need to. It's the medium status ones, who are afraid of becoming outcasts themselves, that do the worst bullying.) Therefore, to raise your own status in a group with ill-defined status, find someone to attack, attack them, and get others to support you in the attack.
But precisely because we never lacked for enemies, it would have been all the more important not to weaken the group by creating internal division.
The group weakening effect of picking on an arbitrary group member is clear and direct. Why would picking on a group member evolve as an indirect way of creating cohesion when the immediate direct effect is to reduce cohesion? Maybe there is some weird feedback mechanism a la runaway selection, but those kinds of arguments are not easy to make out.