The Judge With No Face
Meta's AI moderation is more accurate than humans on average — but at billions of users, that still erases thousands of accounts with no one left to appeal to.
Meta says its AI content-moderation system makes 13% fewer errors and catches 10% more policy violations than the human reviewers it is replacing. That may well be true. And it changes almost nothing about the experience of waking up to find that the account holding fifteen years of your life has simply ceased to exist, with no one to ask why.
Consider the small horror of the notification. You open Instagram to post a photo of your kid's birthday and instead there is a gray screen: your account has been disabled for violating community standards. No specifics. No named human on the other end. You appeal, and the appeal is reviewed — by the same kind of system that erased you — and denied in the time it takes to boil an egg. According to a New York Times report cited across the industry, Meta is leaning harder into exactly this machinery, and a number of Instagram and Facebook users say it has wrongly deleted them.
What does 'better than humans' actually mean at Meta's scale?
Here is the arithmetic that the 13% figure quietly hides. A percentage is a comfort designed for the person reading the dashboard, not the person on the wrong side of the decision. Meta moderates content for billions of users. Trim the error rate by 13% and you still generate a staggering absolute number of wrong calls every single day — wrong calls that, in aggregate, are more accurate than before and, one household at a time, are just as devastating as they ever were. Being more right on average is not the same as being right about you.
The deeper shift is not accuracy. It is appeal. A human moderator, however overworked and underpaid, is a person you can in principle reach — a fellow member of the species who might read your two-line plea and feel the flicker of doubt that says *maybe this one is a mistake*. That flicker is inefficient. It is also the entire moral content of judgment. When you optimize it out, you have not built a better judge. You have built a faster verdict.
What we lose when the verdict has no face
I keep thinking about the word *disabled*. Not deleted, not suspended — disabled, as though your presence were a malfunction that has been corrected. For a small-business owner whose customers live inside a Facebook page, for the diaspora family whose only thread to a grandmother runs through Instagram, for the freelancer whose portfolio is a feed, the account is not a hobby. It is infrastructure. And infrastructure that can vanish on the say-so of a system with no face to appeal to is not really yours. You are a tenant who can be evicted by an algorithm that does not know it is evicting anyone.
The companies will point out, fairly, that human moderation was never a warm human process either — it was a burned-out contractor in a distant office clicking through trauma at inhuman speed. That is true, and it is not a defense of automation so much as an indictment of what we asked people to do before. The honest question is not human versus machine. It is whether accuracy, measured in the aggregate, has quietly become the only value we are still willing to name — and whether the right to be heard by someone who can be surprised by your case is a cost we have decided to stop paying.
An extra ten percent of violations caught is a real good; harmful content ruins real lives too, and that side of the ledger is not nothing. But a moderation system earns legitimacy the way any court does — not by being right most of the time, but by giving the wrongly accused a door they can knock on. Meta is building a better verdict. What it owes the people it erases by mistake is a better appeal, and a face on the other side of it.