Is there a social media bias?

Uneven enforcement reflects uneven rule-breaking, and the evidence offered is anecdotal

Question: Is there a social media bias?
Position‹2 of 2: No, there is not a social media bias
Argument‹2 of 2: Uneven enforcement reflects uneven rule-breaking, and the evidence offered is anecdotal

The argument

Opponents of the bias claim argue that the case for it is built on a selection of examples rather than on any measure of the whole. A platform handling an enormous volume of content will produce a large number of mistaken decisions in absolute terms even if it is accurate in the great majority of cases, and those mistakes can be collected into a compelling list by anyone looking for them. Opponents point out that such lists are assembled by interested parties, that the counterpart list from the opposing side is equally long, and that neither tells you the rate at which either group is affected. The second part of the argument concerns the inference from unequal outcomes. Opponents hold that enforcement falling more heavily on one group is only evidence of bias if that group was breaking the rules at the same rate as everyone else, and that this is assumed rather than shown. If a particular community produces more of the content that violates a given policy — abuse, health misinformation, coordinated inauthentic behaviour — then even-handed enforcement will remove more of its posts. Treating disparate impact as proof of discrimination, opponents argue, requires ignoring the possibility that the underlying conduct differs. Opponents add that the individual cases, once examined, usually have ordinary explanations: an automated classifier error later reversed, a genuine rule breach the poster did not mention, or a decision applied equally to similar posts that went unnoticed. The appearance of a pattern comes from the selection, not the data. Because the case for bias rests on anecdotes gathered by interested parties and on treating uneven outcomes as proof regardless of uneven conduct, this argument concludes, no systematic social media bias has been demonstrated.

Premises

[P1]At the scale platforms operate, even a low error rate yields many mistaken decisions, which can be collected into a persuasive list by any interested party. [P2] Comparable lists exist for every political group, so such collections show nothing about the rate at which each is affected. [P3] Enforcement falling unevenly is evidence of bias only if the underlying rate of rule-breaking is equal, which is assumed rather than demonstrated. [C] Therefore the evidence offered does not establish a systematic social media bias.

Counter-arguments

Proponents of the bias claim reply that the appeal to differing rates of rule-breaking assumes the rules themselves are neutral, when the definitions of harassment, misinformation and hate speech are contestable and are written by the same interested party that enforces them. They add that systematic evidence is unavailable only because the platforms withhold the data and permit no independent audit, so the reliance on anecdote is a consequence of the opacity rather than a weakness of the case.

Rejecting the premises

[Rejecting P1] Error at scale is granted and it explains the existence of a list. It does not explain a direction, and the premise would need those errors to be distributed evenly in order to rule out the bias alleged. [Rejecting P2] That comparable lists exist for every political group shows that each can assemble examples, not that the rates are equal. The premise moves from the availability of anecdote on all sides to the absence of any difference between them. [Rejecting P3] This objection holds, and it cuts both ways: the underlying rate of rule-breaking is equally unmeasured on this side of the dispute, so the premise establishes that neither party has made its case rather than that there is no bias.