The email arrives at 3:12 a.m., as these emails do: your account has been disabled. Tonight it lands with a woman who sells ceramics through Instagram — twelve years of photographs, forty thousand followers, a wholesale order half-negotiated in direct messages, all of it suddenly behind a login screen that no longer recognizes her. The email offers one mercy, styled as a button: she can appeal. So she writes a short, careful, polite paragraph, the kind you write when you believe a person will read it. She explains what her business is, and that she has never done anything resembling the thing she stands accused of. She submits it into a small form, and waits.
What she has actually done is send a second input to the machine that banned her. There is no clerk on the other side of the form, no file being opened, no judge of any description. Her paragraph will be scanned, classified, and in most cases weighed by the same system that reached the first conclusion — a system that cannot tell her what she did, because telling her would teach someone else how to do it. The appeal looks like the start of a process. It is closer to a ritual: a small ceremony of being heard, staged by infrastructure that is not, in any meaningful sense, listening.
Meanwhile, account termination has quietly become a consequential civic event. Moderation used to mean a deleted post; increasingly it means the deletion of a place. A terminated seller loses her storefront, her customer list, her archive. A terminated creator loses an audience assembled over years, and the records that prove it ever existed. These are losses that used to require an eviction notice, a bankruptcy, a court order — events with reasons attached and some possibility of reply. Platforms rebuilt all of that inside private systems, then began deciding it at a scale that forbids individual attention.
The standard defense is that the scale is the point: billions of users, enforcement actions in the hundreds of millions each quarter, human reviewers whose attention is rationed in seconds. All true. But notice the shape of the argument. A company builds a city the size of a civilization, staffs its courthouse like a call center, and then cites the population as the reason due process is impractical. Scale is not weather; it is the business model. And it does double duty — first as the source of the platform’s power over livelihoods, then as the excuse for wielding that power without any of adjudication’s constraints: no hearing, no evidence, no reasons.
The reasons point deserves a pause, because the refusal to explain is not laziness; it is doctrine. Platforms run enforcement like a security operation, and in security you do not publish why the alarm went off, because the write-up becomes a field guide for the next intruder. That is a sane way to fight spam. It has been quietly extended to everyone, so the ceramics seller is processed as a bot that has not confessed yet. But courts operate amid gaming too — every legal system ever built has been gamed — and they publish reasons anyway, because reasons are the only thing that makes power checkable. A decision you cannot interrogate cannot be wrong in any way that matters. It can only be final.
The redress lottery
Here is the awkward part: a working appeals process does exist on every major platform. It just isn’t the form. It is the viral post about being banned, the journalist who emails the press office, the friend of a friend on the inside. Ask anyone who has recovered an account how it happened and the answer is almost never “the appeal worked.” Redress is allocated by attention, which means it flows to the already visible — a justice system run like a trending page.
The formal channel, meanwhile, shows up in the transparency data as statistical noise. Meta’s Community Standards Enforcement Reports are admirably blunt about the ratio: restorations on appeal run in the low single-digit millions per quarter, against enforcement actions in the hundreds of millions, most of it flagged by automation before any human complained. People appeal a small fraction of what gets taken down; a small fraction of those appeals restore anything. The button is real. The ceremony is the product.
Making the machine explain itself
The law has begun to notice the gap between the costume and the court. The EU’s Platform-to-Business Regulation, in force since 2020, already obliges platforms to give business users reasons and, in most cases, thirty days’ notice before termination — an early admission that deleting an account resembles an eviction more than a deleted comment. The Digital Services Act, fully applicable since February 2024, goes further. Enforcement decisions must come with a statement of reasons; appeals must be available free of charge for at least six months, with complaint decisions not left solely to automated means; and users can take disputes to certified out-of-court bodies whose rulings bind the platform. The statements of reasons now flow by the hundreds of millions into a public EU database — the machine explaining itself at machine scale, in a voice no human will ever read, which is its own kind of answer.
In the United States, the same attempt went sideways: Texas and Florida passed laws mandating transparency and appeal procedures, the platforms sued, and in July 2024 the Supreme Court sent the cases back down in Moody v. NetChoice without deciding whether any of it can stand. The direction of travel is set in Europe and contested in America, but the underlying question is identical: whether a compelled reason can be more than boilerplate — whether a statement of reasons is an explanation, or just a longer form of silence.
Days after her paragraph disappears into the form, the ceramics seller gets her answer. The reply confirms, in the passive voice common to all such replies, that her account has been reviewed and found to violate the same community standards it was accused of violating the first time. The accusation has been read back to her in the first person plural, and that is the verdict. What regulators are now testing, for the first time, is whether it can be made otherwise — whether a platform can be forced to say why, in terms specific enough to argue with. Because a reason you can argue with is the entire difference between being governed and being processed. And at the scale these machines now operate, processing is what we have.