Fifth, and closely related to the above point, moderation systems need to be designed to lighten the load on moderators, whether they are admins or the regular users themselves. Paul Graham has become a believer in the "broken windows" theory of maintaining order: small violations of the spirit of expected behavior, if persistent and unchecked, can undermine broader adherence to those norms. To wit: if a rule is unenforced and constantly violated, is it really a rule? The solution that xkcd's Randall Munroe hit upon after reviewing the standard options faced by all rapidly scaling communities--restricted entry, moderators, user moderation, and sub-communities--was a system of passive moderation. Moderation would be automatically applied according to a predetermined set of criteria specifying what qualities a good comment would have. In Munroe's case, originality was the key, and any commenters attempting to say something that had already been said before would be penalized by increasing mute times. A similar project, the StupidFilter, being developed by one of our own, uses Bayesian logic to identify stupid comments based on a seed group of human-identified stupid comments. The criteria for stupidity include: over- or under-capitalization, too many text message abbreviations, excessive use of 'LOL' or exclamation points, and so on. Spam identification systems for email and blog comments (e.g. Akismet for WordPress) do much the same thing, identifying commonalities in junk messages and containing them in a junk/spam purgatory awaiting moderation.

Passive moderation can help solve the problem of moderator overload, just as spam filters aid managing one's email inbox or blog comments. Reducing the number of full-time admins to do moderation reduces the proclivities toward the "iron law of bureaucracy" and toward user-moderation abuse. Like the above passive systems, a Robot9000++ could be set to identify general characteristics of comments that make them good or bad: not only originality, but also ideal length of the post (with diminishing returns after a certain point), presence of links, paragraph structure, and so on. Likewise, it could identify posts that the typical profile of destructive or idiotic behavior: one-liners, ad hominems, common insults, links to shock sites, etc. False positives would be an issue, hopefully less of one over time if it had a Bayesian capacity to learn. But effective admin intervention and/or user moderation could correct erroneously downmodded comments.