The Human in the Loop

The one decision the machine never gets to make — and why "remove the human" is the most dangerous idea in the room. The Vibe Coding Guide.

Somewhere in the system there is a moment the machine always brings to me and never decides on its own. It isn't the hardest moment, or the most technical. Usually it's the one where a real person is on the other end — a customer who's upset, a refund that isn't clean, a call where the rule and the right thing point in slightly different directions. The machine could decide it. It would decide it fast, and consistently, and it would be wrong in a way nobody noticed until it had already happened ten thousand times. So I keep that moment for myself. Not because I'm better at it. Because I'm the one who can be held responsible for it.

That's the part the whole industry is racing past, and it frightens me, because I come from the one place that learned this lesson with real money and occasionally real blood. Everyone is sprinting toward autonomy — take the human out, let it run, watch it scale. And on the easy days it works beautifully; the human was only friction, a slow wet bottleneck in a fast dry machine. The trouble is that the easy days were never the ones the human was there for. A human in the loop was never about being faster or smarter than the machine — we aren't, and pretending otherwise is exactly how you talk yourself into removing us. It was always about something the machine structurally cannot supply: someone who is accountable. Someone with skin in the game. Someone the consequence actually lands on.

You cannot hold a machine responsible. You can't fire it, can't sue it, can't sit it across a table from the person it hurt, can't make it lie awake. When a system removes the last human and then does something catastrophic and irreversible at machine speed — the trade that cascades, the bug that ships to everyone at once, the confident decision that is confidently wrong — there is no one who saw it coming and no one to answer for it afterward. The accountability didn't move to the machine. It evaporated. And a decision no one is accountable for is a decision no one is really making; it's just happening, to people, at scale. That's the most dangerous thing I know how to build, and the industry is building it on purpose, because it's cheaper.

So the line I hold is simple to say and hard to keep: automate the toil, never the accountability. Take the human out of the fetching and the routing and the thousand repetitions — the wires, the parts that were always beneath a person anyway. But the decisions that touch a real life, that can't be undone, that someone will one day have to stand up and answer for — those keep a human at the gate, with real authority, not as a courtesy. There's a name for the ladder this implies, the delegation continuum: at one end the machine merely informs you, at the far end it acts entirely alone, and the whole craft is putting each decision on the right rung. Most things belong far toward the autonomous end — that is the point of everything I've built. The few that don't, the irreversible and the human, stay near my end of the ladder, and they stay there on purpose.

In my own company those gates are concrete, and a little obsessive. When the machine proposes a new piece of work — a task it thinks a department should pick up — it doesn't get to act on it. It surfaces the idea and stops, and waits for me to approve it; only then does it go off and flesh the thing out; and then it stops again, and waits for me to confirm the fleshed-out version before a single human downstream ever sees it. Two gates, both me, on either side of the machine's best thinking. Every outbound email the system drafts sits in a queue marked pending until a person releases it — nothing reaches a real inbox on the machine's say-so alone. It looks like friction, and it is friction, and the friction is the point: it's the seam where a human stays answerable for what the company says and does. I automated the work of getting to the decision. I did not automate the decision — and I built the thing so that it can't quietly start making those for me.

Now the warning, and it's a subtle one, because you can get this wrong while looking like you got it right. A human in the loop is worthless if they're only there to absorb the blame. There's an ugly pattern researchers named the moral crumple zone: a company keeps a person at the controls not so they can genuinely steer, but so there's somebody to point at when the machine fails. The autopilot does everything, the human can't realistically intervene in the half-second that matters, and when it crashes the human takes the blame the system was designed to shed. That isn't a human in the loop. That's a scapegoat with a job title. The real thing means the human has actual authority and actual understanding at the gate — enough context to say no, enough power to make the no stick, enough time to mean it. If your human can't truly stop the machine, you don't have a human in the loop. You have a liability sponge.

Strip it all the way down and there's a small, stubborn core that stays human no matter how good the machine gets. What is worth doing at all — the machine can optimize the how, but it cannot want the what. What we refuse to do even though we could — the values, the lines we hold for no profitable reason. Who answers when it goes wrong — the accountability that has to live in a person. And the genuinely new — the situation no rule covers and no past case fits, where you aren't retrieving an answer but making one. Everything else, I hand to the machine with my blessing. These four I keep. Not out of vanity. Because a company, a system, a society that hands these to a machine has quietly decided that nobody is home.

People treat keeping the human in the loop as a concession — a temporary inefficiency we'll engineer away once the models get good enough. I've come to believe the reverse. The better the machine gets, the more it matters that a human stays exactly where the consequences land, because the machine's rising competence is precisely what tempts everyone to stop watching. I keep my hand on the gate for one reason: I'm the one who has to be able to look the person on the other end in the eye. That was always the job. It's the only part of it that was ever really mine.

Next: the distance between us — what every last piece of this was actually for.