The platforms' own rules are enforced unevenly

The argument

This argument focuses on the gap between the rules as written and the rules as applied. Proponents point to the familiar pattern in which comparable posts meet opposite fates: one account is suspended for wording that another uses without consequence, a claim is labelled when made by one side of a dispute and left alone when made by the other, and the same image is removed in one context and permitted in another. Where a published rule produces inconsistent outcomes across similar cases, proponents argue, something other than the rule is deciding. Proponents identify the mechanisms that produce the unevenness. Enforcement is largely triggered by reporting, so the better-organised and more numerous complainants get more enforcement directed at their opponents. Prominent accounts are routinely handled by exception rather than by the ordinary process, with decisions escalated and sometimes reversed on the basis of their commercial or political importance. Automated classifiers are trained on past decisions and so reproduce whatever slant those decisions contained, at scale and invisibly. None of these is a deliberate conspiracy, and proponents argue none needs to be: each produces systematic asymmetry on its own. Proponents hold that the appeals process confirms rather than corrects the problem. Appeals are decided by the same company, with no external review, no published reasoning and no remedy for the reach already lost. A rule that is applied inconsistently and reviewed only by its author is not a rule constraining the platform but an instrument available to it. Because the platforms' stated policies are enforced unevenly through reporting dynamics, exceptions for prominent accounts and classifiers trained on prior decisions, this argument concludes, there is a social media bias.

Premises

Counter-arguments

Opponents of the bias claim reply that uneven enforcement is evidence of bias only if rule-breaking is evenly distributed, which is assumed rather than shown; where one community produces more of the content a policy prohibits, even-handed enforcement removes more of its posts. They add that at this scale a low error rate still yields many wrong decisions, which any interested party can collect into a persuasive list, and that comparable lists exist for every political group.

Rejecting the premises

[Rejecting P1] That comparable posts and accounts receive different treatment is asserted, and what makes two of them comparable is precisely what the dispute is about. At the volumes involved, differing outcomes are also what a low error rate produces. [Rejecting P2] Reporting patterns, special handling of prominent accounts and classifiers trained on past decisions do produce asymmetry, and asymmetry by those routes is not viewpoint bias: it tracks which users report, which accounts are prominent and what the training data contained. [Rejecting C] The conclusion requires a direction where the premises establish inconsistency. Uneven enforcement with no account of whom it disadvantages supports a finding of arbitrariness rather than of bias.