The author compares small communities to the iterated prisoner's dilemma, which has been 'solved,' with the best strategy being 'tit-for-tat.' Iterated versions of coordination games rely on reputation being known by all players. As communities scale, the number of players rises, and the incentive to coordinate falls because it's more difficult to enforce reputational costs to bad behavior.
Except online communities do have reputation and rewards/costs associated with behavior, usually in the form of points or "karma" or some other distinguishing feature. But, even in small communities, reputation systems do not prevent trolling, flaming, or crapflooding (i.e. 'defection'), and community voting is a notoriously bad system for establishing the true quality of interaction/contribution. Moderators who can ban defectors are still necessary, because reputation systems aren't enough.
Further, the prisoner's dilemma model fails to describe his prior example: the awful YouTube post. This is not a 'defection' in the way that trolling or flaming is. It is either an actual idiot or a foreign language speaker writing broken English. It's not a malicious post, even if it is useless noise. Scale will bring in anti-community trolls, but it will also bring in idiots, teenagers, and people otherwise unable to contribute at a high level (foreign language speaker, people posting from their phone, etc). Their earnest but shitty contributions aren't a defection and the iterated prisoner's dilemma reputational model does not take into account their effect on the signal-to-noise ratio and community decay.