Camille Hanson was asleep in Portugal when the machine decided she was a fraud.
The English teacher, 40, woke one night in March to a Meta notification: her Facebook and Instagram accounts, the storefront for the business she runs with her husband and its nearly one million followers, had been flagged for deletion. The stated reason was a violation of Meta's rules on fraud and deception. She appealed. A week later the answer came back — automated, absolute. "All your information will be permanently deleted," the message read. "You cannot request another review of this decision."
"I'm just an English teacher," Hanson told The New York Times, which reported her case as part of a wave of wrongful account removals sweeping across creators and small businesses in the United States, Australia and New Zealand. She got her account back only after the Times asked Meta about it.
That sequence — catastrophic ban, robotic appeal, restoration only when a journalist intervenes — sits at the center of a widening dispute over how Meta polices its platforms. And it collides head-on with the company's central defense of its automated systems, made this week: that the software is simply better at the job than people are.
The numbers Meta is leaning on
Meta says its newer AI moderation tools make roughly 13 percent fewer mistakes than human reviewers and catch about 10 percent more policy violations, a claim reported by the Times and echoed across coverage of the company's enforcement overhaul. The company has been steadily replacing human reviewers with automated systems since Chief Executive Mark Zuckerberg's early-2025 pivot toward looser moderation and fewer "false flags," a shift that Meta has credited with cutting censorship errors sharply.
Confronted with the deleted accounts, Meta drew a line between old and new. The accounts examined by reporters, the company said, were banned by older moderation systems — not its latest AI tools. In other words: the errors users are living through belong to yesterday's technology, and the improved figures describe tomorrow's.
For affected users, the distinction is cold comfort. The bans arrive with vague or alarming justifications — some users were accused of the single most serious violation in Meta's rulebook, child sexual exploitation — and the door back is guarded by the same kind of automation that shut it. In documented cases, appeals were reviewed by AI, rejected within hours, and marked final. More than 60,000 people have signed a petition demanding that Meta explain its bans and let humans review appeals.
"Meta handed account moderation to AI and laid off many of the humans who used to check it," is how the pattern was summarized in reporting on the bans. The Oversight Board, Meta's own quasi-independent review body, found this year that the company's permanent account bans suffered from due-process and transparency failures, and it has pressed Meta to be clearer about how and why accounts disappear.
Meta, for its part, maintains that mistakes are rare relative to the volume it handles, that it is investing in better detection, and that the overwhelming majority of enforcement actions are correct.
Why it matters
The fight here is not really about whether Meta's AI is more accurate on average. It may well be. The fight is about what "more accurate on average" means when the denominator is measured in billions.
Meta moderates content and accounts across a user base of roughly three billion people, taking action on enormous volumes of posts and profiles. At that scale, a system that is 13 percent better than humans still generates a staggering absolute number of wrong decisions. Shave the error rate and you still leave, plausibly, hundreds of thousands or millions of mistaken actions — each one an aggregate rounding error and, for the person on the receiving end, a small catastrophe. A makeup artist loses 48,000 followers. A disability advocate is silenced. A Juneteenth nonprofit vanishes. An English teacher is branded a fraud in front of a million people.
This is the core tension of automated enforcement: aggregate accuracy and individual harm are different measurements, and improving the first does nothing, on its own, to remedy the second. Human moderators were never just more or less accurate than a model; they were also the mechanism by which a wrong call could be caught and reversed by someone with the authority and empathy to say "this is obviously a mistake." Automate the enforcement and automate the appeal, and you remove not just the labor cost but the escape hatch.
That is the missing piece in Meta's math. The company can cite a better error rate, but it has not shown a scaled human appeals mechanism to match the scale of its automated judgments. When the only reliable way to reach a person is to know a reporter, due process has effectively been privatized to those with press connections — which is another way of saying it does not exist for almost everyone.
What to watch
Three things will signal whether Meta's numbers translate into fairness or just efficiency. First, the appeals architecture: does Meta build a genuine, accessible human review path — the thing the Oversight Board and 60,000 petitioners are demanding — or does it keep routing appeals back through the same models that issued the bans? Second, transparency: whether Meta discloses raw error volumes, not just relative improvement percentages, so outsiders can judge how many people its "better" system still wrongs. And third, regulation: European regulators operating under the Digital Services Act, along with lawmakers scrutinizing automated decision-making, may force disclosures and appeal rights that Meta has so far offered voluntarily and unevenly. Until then, the gap between Meta's statistics and its users' experience will keep being bridged, one at a time, by phone calls from journalists.
“I'm just an English teacher.”— Camille Hanson, Instagram creator wrongly banned