Passage A:

Content-moderation systems built on machine-learning classifiers now review a substantial majority of the billions of posts submitted daily to large platforms, and recent comparative audits find that these systems flag policy-violating content at an accuracy rate matching or exceeding that of trained human moderators reviewing the same sample of posts. Human moderators, by contrast, exhibit substantial rater-to-rater variation: two moderators presented with an identical post disagree about whether it violates policy far more often than two independent runs of the same classifier disagree with each other. Given the sheer volume of content requiring review, a platform relying primarily on human moderators would need to either slow removal dramatically or tolerate wide inconsistency in which posts are removed. Automated systems resolve this tension by delivering both the necessary speed and a level of internal consistency human review at this scale cannot match.

Passage B:

Suppose the audits are exactly right: a classifier agrees with itself far more reliably than two human moderators agree with each other. That fact alone does not tell us whether the classifier's consistent judgment is the correct one. A rule applied uniformly is only a virtue if the rule is a good one; applied uniformly, a flawed classification standard produces flawed removals at scale, with no moderator positioned to recognize that a given post is the unusual case the general rule gets wrong. Human inconsistency, whatever its costs, has the compensating feature that an individual moderator can depart from a rule of thumb when context plainly calls for it—a capacity no fixed classifier possesses, however consistently it applies its own, uninspectable standard.