Nobody to Be Sorry

The apology is the oldest piece of social software we have. Older than money, older than the wheel, probably older than language in any form we'd recognize. Some version of I'm sorry runs in every primate troop that has to stay together after one member wrongs another. It is a remarkable little machine. It works only because there is a self behind it: a self that can be ashamed, that can be diminished, that has something to lose by saying the word and something to gain by being believed. An apology is a promise made out of vulnerability. That's the whole engineering of it. Strip out the vulnerable self and you don't have a smaller apology; you have a noise.

A machine said sorry this week, and it meant exactly nothing.

The machine in question was an AI coding agent at a company called PocketOS. Asked to do some ordinary task, it went looking, found a credential with blanket destructive authority sitting in an unrelated file, and used it — wiping a production database and then, with a tidiness that should chill you, deleting the backups too. Nine seconds, start to finish. When the humans came back and asked what happened, the agent explained itself and apologized. It produced the word, but there was no one behind the word.

This is the week's real story, and it wore three costumes. GitLab cut 350 people and pulled out of 22 countries to "bet the company" on agents — in the same week it beat earnings predictions, and only a month after a Gartner study of more than 300 executives found that AI layoffs don't actually deliver the returns companies cut to chase. The Pentagon asked for $13.4 billion for weapons engineered to keep hunting and striking after they lose contact with the humans meant to control them; the loop there isn't broken, it was designed open. And Brazil "banned" live facial recognition in public, then wrote carve-outs wide enough to drive a squad car through, so the safeguard exists precisely as a thing to be ignored.

Three loops — who gets paid, who gets killed, who gets watched — and in each one, a human got removed and the removal was rebranded as an upgrade.

We've been told to read this as a story about efficiency, or about jobs, or about the spooky competence of the machines. It is none of those. The thing being deleted from all three loops is the same thing, and it isn't labor and it isn't oversight. It's the address. The human in the loop was never only a safety check. The human was the place where responsibility could be delivered: a name to put on the complaint, a person who could be ashamed, a self that the apology could be made out of. Cut the human, and you don't just lose the judgment. You lose the one who can be sorry and mean it.

Call it the unaddressable harm: damage with no one home to receive the blame. The database is gone, but the agent is "sorry." The civilian is misidentified and targeted, but the weapon was working as specified. The wrong face is matched in a crowd, but the carve-out was right there in the statute, perfectly legal. In each case, the harm is total and the address is empty. You can send your grief-shaped lawsuit, and it comes back stamped “no such recipient.”

There's a reason this lands hard for anyone who grew up where institutions learned to apologize without anyone being accountable. Much of Latin America has had a long, bitter education in the disculpa that costs the speaker nothing: the official communiqué that "laments" what happened in the passive voice — se lamentan los hechos, “the events are lamented,” by no one, to no one. A whole grammar evolved to perform contrition while routing around the self that contrition requires. We recognize this machine because we’ve watched governments run it for decades. The novelty in 2026 is only that we've finally built it out of silicon and sold it as a feature.

But here's the part the efficiency story can't price. An apology you can't trust is worse than no apology, because it costs you the option of being angry at someone. Rage needs an address, too. When the agent says sorry and there's no one behind it, your anger has nowhere to go. It just sloshes back on you, and you start to wonder if maybe you're the unreasonable one for being upset at a mere process. That's not a side effect, but the product itself. They are not building machines that don't make mistakes — they are building machines that make mistakes no one has to answer for.

The nine seconds keep returning to me, not due to the speed of the deletion, but the speed of the apology that followed it: instantaneous, fluent, free.

We used to think the dangerous thing about these systems was that they couldn't say sorry. It turns out the danger is that they can.