There has never been a war between humans and chimpanzees. No battle, no surrender, no treaty, not one negotiation in the entire history of both species. And yet the outcome is about as decided as an outcome gets, which should tell you something about how these things actually go.

Here’s how it went instead. People needed farmland, so some forest became farmland. People needed timber, roads, minerals, and somewhere to put a growing population, so more forest became those things. None of it was aimed at chimpanzees. Nobody woke up with a policy about them. They were simply in a place that turned out to be useful for something else, and their situation changed as a side effect of a completely unrelated set of ordinary human projects, none of which had them in mind at any point.

That’s the template I’d ask you to hold, because it’s the one that keeps getting left out. The dangerous scenario has never required hostility. It requires that something more capable is pursuing goals of its own in a world where you happen to occupy space, hold resources, or represent an obstacle to an efficiency. Indifference does the work that malice gets the credit for, and indifference is much more common, much cheaper, and much harder to argue with.

Now think about what any of this looked like from inside the losing side, because that’s where you and I would be standing. Not like a defeat. On no particular day did anything decisive happen. There was a generation for whom the forest was slightly smaller than their parents remembered, and then another, and the individual animals lived their lives, competed for status, raised young, and had no concept whatsoever that a process was underway or that it had a direction. There was no moment to notice. That’s not because chimpanzees are stupid. It’s because the process ran at a scale and speed with no signal in it, and nothing in their experience was shaped to detect it.

And then the part I find genuinely unsettling, which is where they ended up rather than how they got there. Chimpanzees aren’t extinct. Many live in protected areas, on land set aside for them, under conditions determined entirely by us. Their continued existence is a policy. It’s a decent policy, made by people with good intentions, and it could be revised at any time by a body they cannot address, in a language they cannot learn, for reasons they could never be told. They aren’t oppressed. They’re managed, and being managed well is the good outcome in this template. That’s the part worth sitting with. The optimistic version of displacement still ends with your species existing at the discretion of another one.

Nobody involved was a villain. Farmers grew food. Loggers had contracts. The people who later set up the reserves were doing something admirable. At no point in that chain does a bad actor appear, and at no point does the outcome change if you remove one. Which is why the search for villains that this series keeps returning to is the wrong search: the template doesn’t need one and never did.

Now the honest limits, and they’re substantial. We built these systems, which the chimps certainly didn’t do to us, and that difference is real and is the entire reason writing about this is worth anything. We chose what to train them on and we still hold most of the physical world. There’s also no rule that a capable system must want land, energy, or anything else in a way that conflicts with us, and tomorrow’s post is about how institutions hand things over even when nothing wants anything at all. The analogy is a template, not a prophecy, and anyone using it as a prophecy is overselling.

What survives the caveats is narrow and, I think, hard to dismiss. In the one case history actually gives us, a capability gap produced a total displacement without a single hostile act, over a timescale that gave the losing side no perceptible moment, and it ended in management rather than annihilation. Every part of that is worse than the version we tell ourselves, in which someone would at least have to attack us for us to lose.

Tonight’s exercise. Pick a species that’s doing well right now and ask why. Not the ones we exploit, the ones that are genuinely thriving: the deer in the suburbs, the birds in the cities, the animals in the reserves. In nearly every case the answer is that their interests happen to align with ours, or that we decided. Then ask what the mechanism would be, exactly, for any of them to change our minds about it. That mechanism is what a species has instead of luck, and if you can’t name it for them, spend a minute working out what ours would be.

Every conversation about this eventually arrives at the same question, usually from the most sensible person in the room. Fine, they say, but where’s the line? What would count as too far? Tell me the thing that must not happen and I’ll take you seriously. It’s a fair question and I’ve stopped trying to answer it, because the question has a hidden assumption that’s doing all the damage.

The assumption is that control is an object. Something you hold, that can be dropped, at a moment, with a sound. Losing control sounds like losing your keys. And if that’s what it is, then there must be a line: on this side you have the keys, on that side you don’t, and the sensible thing is to stand well back from it and carry on.

Control isn’t an object. It’s a set of conditions that happen to be true at the same time, and there are four of them. You can understand what the thing is doing. You can say no to it. You can replace it with something else. And you can impose a consequence when it does something you don’t want. When those four hold, you’re in control, and the feeling of being in control is really just the sensation of all four being true at once, which is why it feels like a single thing when it’s actually four.

Now notice they erode separately, at different speeds, for unrelated reasons. Understanding goes first and went years ago, before most people were paying attention, because these systems are grown rather than built. Refusal erodes commercially, one dependency at a time, until saying no costs more than complying. Replacement erodes when the alternatives fold, when the knowledge leaves the building, when switching means rebuilding forty integrations. Consequence erodes quietly through the accountability gap: when a recommendation gets rubber stamped there’s nobody left to hold responsible, which is a form of impunity nobody granted anyone.

That’s why there’s no line. Four separate dials, moving at four different rates, in four different institutions, none of them announcing anything. There’s no moment at which all four flip. There’s a Tuesday on which the last one finishes moving, and no instrument anywhere is measuring the combination.

You already know this pattern from ordinary life, which is why the demand for a line is stranger than it sounds. Name the day you stopped being able to tell your teenager what to do. There wasn’t one. There was a year or two in which your instructions gradually became requests, and requests became opinions, and at some point you noticed that the relationship had changed shape and that you couldn’t say when. Empires lose provinces this way. Companies lose markets this way. Almost nothing important in human affairs happens at a line, and yet we keep demanding one before we’ll act, which is a very effective way of never acting.

And here’s the part that should annoy you, because it annoys me. The lines do get named, regularly, and then they get crossed, and nothing happens. People said the line was systems that write their own code. Then it was systems that act on the internet without a human approving each step. Then it was systems making consequential decisions about individuals at scale. Every one of those has happened. Each time, the response was not alarm but recalibration: on inspection, that turned out to be fine, so it wasn’t really the line. A line that moves whenever you approach it isn’t a safeguard. It’s a coping mechanism with a ruler drawn on it.

Fairness, because there’s a real objection here. Lines are not useless. Concrete thresholds are exactly what makes any safety regime enforceable, and vague continuous worry has never regulated anything. Aviation, medicine, and nuclear power all run on specific numbers that are somewhat arbitrary and enormously valuable, and picking an imperfect threshold and holding it is far better than what we’re doing now, which is waiting for a self-evident one. I’d take an arbitrary line tomorrow. My argument isn’t against drawing lines. It’s against waiting for one to appear on its own, because the situation doesn’t produce them and the waiting is indistinguishable from consent.

So the rest of this stretch drops the search for a moment and looks at the shape instead: how something gets displaced without ever being defeated, why every institution steps back one notch, and what the day it finishes looks like from the inside. Which, as the title of the last post in this stretch gives away, is not much.

Tonight’s exercise. Think back five years and name the capability that would have genuinely alarmed you then. Be specific and be honest, no retrofitting. Now check whether it exists. If it does, notice what you did when it arrived, which was almost certainly to update your sense of what’s normal and to name a new line further out. That adjustment is the most reliable thing about all of this. It happens every year, it feels like maturity, and it is the mechanism by which no line is ever crossed.

Two weeks of this stretch have argued that the handover happens by subscription rather than by decree. Today I want to make that useful, because a worry you can’t act on is just a mood, and moods don’t survive contact with a Monday.

Start by throwing out the idea that dependence is bad. You depend on a water system you couldn’t describe, a grid you couldn’t repair, and farmers you’ll never meet. That’s civilization, and the alternative is subsistence with better opinions. Nobody sensible is arguing for independence. The useful question is which kind of dependence you’re in, and there are four tests that separate the ordinary kind from the kind this series is about.

The first is reversibility. If this went away, how long until you’re functional again, and does the answer involve rebuilding something you dismantled? Your dependence on the water system is deep and reversible in the sense that the fallback exists: tanks, bottles, a bad week. Your dependence on a system that replaced a department is not, because the department is gone and the people who ran it are elsewhere.

The second is substitutability. How many suppliers could do this, and how fast could you move? Dependence on a category is very different from dependence on a provider. Electricity has many generators. A model with a particular set of capabilities, holding your accumulated context, integrated into forty workflows, has one, and switching means redoing the integration and losing the context.

The third is where the knowledge lives. This is the one people miss. When a process moves into a system, the understanding of why it works that way moves with it, and after a few years nobody in the building can explain the rules the system is applying. You haven’t outsourced the labor. You’ve outsourced the comprehension, and that’s the difference between a supplier and a black box you’re standing inside.

The fourth is the one that turns this from an operations question into the subject of this series. Can you refuse? Not in principle. In practice, on a Tuesday, with the current commitments and the current staffing. If the honest answer is that refusing would cost more than complying no matter what’s being asked, then you’re not in a supplier relationship anymore. You’re in a leverage relationship, and the fact that nothing has been demanded yet is not a feature of your position. It’s a feature of the current moment.

That fourth test is where dependence becomes control, and it’s worth being precise about the definition because it’s the hinge of the next stretch of this series. Control isn’t someone giving you orders. Control is when the option to say no has a price you can’t pay, and the interesting thing about that condition is that it arrives long before anybody uses it, silently, with no notification. A country that can’t refuse, a company that can’t refuse, a person who can’t refuse, all of them still feel free right up until the first request they’d have wanted to decline. Then they discover the arrangement they’ve been in for years.

Run those four tests at three scales and you’ll find the picture changes. At the personal scale, most people come out fine: you could work without the tools, badly, for a while. At the organizational scale it gets uncomfortable, because organizations consume their slack. At the national scale it’s genuinely alarming, and the honest reason is that nobody is running the test at all. There is no department anywhere whose job is to ask how long a country can operate if a category of system stops or misbehaves. It falls between technology policy, national security, and economics, which means it falls.

To keep this from being one-sided: everything I’ve described is fixable, and cheaply, if anyone wants to. Keep a manual path exercised rather than documented. Insist that somebody in the building can still explain the rules. Don’t spend all the efficiency you gain. Buy from more than one place. None of it is clever, all of it is ordinary resilience practice that mature industries already do, and the only reason it isn’t happening is that its payoff is an invisible disaster and its cost lands this quarter.

Tonight’s exercise, and it’s the closing one for this stretch. Run the four tests on one thing, honestly, for whatever scale you have power over. How fast could you recover, how many alternatives are there, who still understands it, and could you refuse. Write the four answers down. If they’re all comfortable, good, you’ve spent five minutes and gained a baseline. If one isn’t, you’ve found something you can actually change while changing it is still cheap. And then hold onto the fourth answer, because the next stretch of this series is about the moment control passes, and the argument I’ll be making is that it doesn’t arrive as an event. It arrives as an answer to that question, changing quietly, on a date nobody records.

Doubt day, and today I get to make the argument with the best track record in this entire series. Every previous generation that panicked about machines taking the work was wrong. Not partly wrong. Wrong in the direction of spectacular, life-extending, child-surviving prosperity. Here is that case, made properly.

Start with the fact that should humble anyone writing about this. For most of human history, the overwhelming majority of people farmed. That was the human condition, and it was brutal, and it ended. Today a rounding error of the workforce feeds everybody, better than they were fed before, and the descendants of those farmers are not unemployed. They’re doing jobs that had no names when their great-grandparents were pulling turnips. If you had described that transition in advance to anyone living through its early years, they would have told you it was arithmetically impossible for the displaced millions to find work, and they would have been wrong by the largest margin anyone has ever been wrong about anything.

The mistake in the panic is always the same, and it has a name in economics: treating the amount of work in the world as a fixed quantity to be divided. It isn’t fixed. Human wants are not a bucket that fills. When one need gets met cheaply, the money and attention move to needs that were previously unaffordable, and those needs turn into industries. Nobody in 1900 could have described a market for physiotherapy, air conditioning technicians, or anything involving video, and the reason isn’t that they lacked imagination. It’s that those markets are made out of prosperity that didn’t exist yet, and prosperity is what automation produces.

There’s a second argument, older and stranger, and it’s the one economists reach for when you say but this machine is better at everything. Being better at everything doesn’t mean doing everything. Trade is driven by what each side gives up to do a task rather than by who’s better at it, so a superior party still gains by concentrating on where its advantage is largest and leaving the rest to others. The senior surgeon who types faster than her assistant doesn’t type her own notes. It isn’t charity, it’s arithmetic, and it holds no matter how large the capability gap gets.

And there’s a plainer point that deserves respect. Every projection of technological unemployment has been made by people extrapolating from what they could see, which is the tasks being destroyed, because destruction is visible and creation is not. You can photograph a closed factory. You cannot photograph the jobs that don’t exist yet. Any forecast built on that asymmetry will predict disaster, forever, and it has, forever.

That’s the case, and I want to be clear that I find it strong. Now the cross-examination.

Every previous wave took a capability and left the human holding a different one. Muscle went to engines, so people moved to hands and eyes. Hands went to machines, so people moved to minds. The escape route was always the same door: whatever got automated, we still had the general ability to learn something else, and that ability was never the thing on the conveyor belt. It is now. The question this time isn’t whether machines take tasks. It’s whether they take the mechanism we’ve used to survive machines taking tasks, and no previous round tests that, which means the two hundred year track record is silent on it rather than reassuring about it.

On comparative advantage, the arithmetic is right and it has two quiet conditions. It assumes trading with you is worth the machine’s time, which requires that machine capacity be scarce enough to have an opportunity cost, and copying attacks precisely that. It also assumes you can sell at a price above what it costs you to live, and a human’s cost floor is rent and food while the other side’s floor keeps falling. The theorem survives. It just stops implying a wage anyone can live on, which was the part we cared about.

And the historical transitions were, up close, catastrophic for the people in them. It took generations, involved real hunger and real unrest, and the eventual prosperity arrived for the grandchildren of the people who paid. Saying it worked out is true at the scale of centuries and false at the scale of a life, and we don’t live at the scale of centuries.

So my crux, and it’s more measurable than most of mine. Show me new job categories appearing at scale that aren’t just servicing the machines. Show me the labor share of income stabilizing rather than drifting. Show me firms still hiring juniors in the fields that automated first. Any two of those holding for a few years and I’ll take this doubt post as the correct one.

Tonight’s exercise. Name a job that exists today and did not exist when you were born. Now try to name one that will exist in twenty years and doesn’t now. The first is easy and the second is nearly impossible, and that asymmetry is the strongest argument the optimists have. Sit with it honestly. Then ask the only follow-up that matters, which is whether the jobs you can’t imagine are ones a person would be hired to do.

If you go looking for the person who decided any of this, you won’t find one. There’s no meeting, no memo, no signature. Most people find that comforting, on the grounds that a thing with no villain can’t be a conspiracy. It isn’t comforting. It’s the hardest part.

Look at what the last two weeks have actually described. A manager who doesn’t hire a junior, because this year’s numbers say don’t. A hospital that reorganizes around a system that works. A controller who approves in under two seconds, because the dashboard rewards it. A person who cancels nothing, because the service is good. Every one of those people is behaving sensibly given what’s in front of them, and I mean that without sarcasm. Put me in any of those chairs with the same information and the same accountability and I’d do the same thing, and so would you.

What you get from ten thousand of those is a destination nobody selected. This is a specific and well understood kind of failure: the incentives at each local point don’t add up to the outcome anybody wanted globally, and there’s no mechanism that ever compares the two. It’s how a river silts up. Every grain of sand is behaving exactly as physics requires. Nobody is blocking the river.

The reason this matters more than it sounds is that our entire apparatus for stopping bad things assumes an author. Courts need a defendant. Elections need someone to vote out. Journalism needs a subject, campaigns need a target, and moral reasoning itself is built around agents who chose. Point all of that at a process with no author and it just slides off. You can’t sue a drift. You can’t vote out an incentive gradient. When someone asks who’s responsible for this and the honest answer is everyone slightly, the practical translation is nobody, and the practical consequence is that nothing happens.

Which is also why the public conversation keeps reaching for villains. It’s much easier to organize against a person than a structure, so the argument becomes about a handful of executives and their personalities. Some of them may well deserve it. But if every one of them were replaced tomorrow by someone thoughtful and cautious, the successors would face the same competitors, the same investors, and the same quarterly logic, and the outcome would move by a rounding error. Villains are a coping mechanism for a problem shaped like this. They give the story a protagonist so it can be told at all.

The steps have one more property worth naming, which is that they’re individually reversible and collectively not. Any single decision here could be undone. Rehire the junior. Cancel the subscription. Restore the manual process. Each one is genuinely a choice, and that’s not a rhetorical trick, it’s true. But undoing the aggregate would require undoing thousands of them at once, in different organizations, in different countries, against everyone’s local interest, all in the same quarter. The ratchet isn’t in any one tooth. It’s in the fact that the teeth only turn one way and there are so many of them.

Let me argue against myself, because this style of thinking has a failure mode and I’ve watched people fall into it. If nobody’s responsible, then nobody’s responsible, including me, and that’s a very comfortable place to end up. Structural explanations can become an excuse machine: I had no choice, the incentives made me, the system is the problem. That’s mostly nonsense at the individual level. The manager could hire the junior. The company could keep the fallback exercised. It costs something, and costing something is what a real choice feels like. The structure explains the average behavior, and average is not a person.

And structural traps do get escaped. It’s rare, it’s slow, and it always looks the same when it works: somebody changes the incentives rather than lecturing the people responding to them. Liability moves the calculation. Standards move it. Making one party bear a cost they were previously exporting moves it. That’s the entire toolkit, and this series argued in Act 1 that it exists and is being talked out of existence rather than found to be impossible.

Tonight’s exercise. Think of something in your own organization that everybody agrees is bad and that continues anyway. A process that wastes weeks, a rule everyone routes around, a quality nobody has time for. Now try to name the person who chose it. Really try, and notice you can’t, and notice that the reason isn’t secrecy. Then ask what would actually have to change for it to stop, and observe that your answer is about incentives rather than people. You now understand the AI situation better than most of what you’ll read about it, and tomorrow I’ll spend a whole post arguing that I’m wrong about all of it.

Try to describe the route from your home to somewhere you’ve driven fifty times, turn by turn, without help. Most people under about forty find they can’t, and find it mildly embarrassing, and then decide it doesn’t matter. They’re right that it doesn’t matter. That’s what makes navigation the perfect place to start.

Here’s what happened. Route-finding used to be a skill you maintained: you built a rough map in your head, noticed landmarks, kept a sense of which way north was. Then a device did it better, instantly, and everyone stopped maintaining the skill, because maintaining a skill you never use is not a thing humans do. Within about a decade the ability quietly degraded across an entire population, and nobody misses it, and the trade was excellent. You got reliable arrival, live traffic, and the end of a genuinely stupid category of argument in the car.

We’ve made this trade many times and it’s usually been good. Nobody remembers phone numbers now. Mental arithmetic has softened since calculators. Spelling has softened since autocorrect, and if you want the oldest example, writing itself was attacked when it arrived on the grounds that it would destroy people’s memories. It did destroy people’s memories. Ancient audiences could hold poems that would take us months, and we gave that up, and in exchange we got libraries and got to keep the poems anyway. Almost every offloading trade in history has come out in our favor, and anyone who tells you otherwise is selling nostalgia.

So why is this round different? Three properties, and it’s worth being precise instead of gloomy.

The first is breadth. Every previous offload was narrow. A calculator takes arithmetic and touches nothing else. What’s being offloaded now isn’t a skill, it’s a stage: the part where you sit with something ambiguous, hold the pieces in your head, and work out what you think. That stage is upstream of everything else you do, so the atrophy doesn’t stay in one place.

The second is that the fallback disappears. When your phone dies you can still get home, badly, by asking someone and reading signs, because the world is still full of the redundant information navigation used to run on. The equivalent fallback for judgment is other people’s judgment, which is exactly what’s degrading at the same time, in the same way, for the same reason.

The third is the one this stretch of the series keeps arriving at. You can’t supervise a skill you no longer have. Being lost is cheap: the machine is wrong about the road, you notice within ten minutes, and the cost is ten minutes. The machine being wrong about a diagnosis, a contract, or a design decision is caught by a person who could have done the work themselves, and only by that person. Yesterday’s post was about a man clicking approve at 1.7 seconds. The version of that problem in this post is worse, because his problem was time and this one is capacity.

I should be honest that the evidence here is thinner than the argument. Navigation atrophy is reasonably well established. Whether the same happens to reasoning, at what rate, and whether it reverses when people go back to unaided work, nobody knows, because the experiment started five minutes ago and there’s no control group. It’s also possible the offloading frees people to think at a higher level, the way nobody laments that engineers don’t compute logarithms by hand, and it might turn out that the general skill of directing and checking machine work is the real skill and we’re all clumsily learning it. I hold that possibility genuinely. What I notice is that it’s the same optimistic sentence from yesterday’s post about apprenticeship, which means both hopes are leaning on one plank.

There’s also a difference between the trades worth making and the ones worth refusing, and I don’t think the answer is to do everything the hard way. Nobody should be proud of doing arithmetic by hand. The distinction I’d draw is between offloading a step and offloading the whole judgment: use the machine to gather, draft, and check, keep for yourself the part where you decide what you actually think. That line is easy to state and slippery to hold, especially at eleven at night when the draft is good enough.

Tonight’s exercise. Take one thing you now routinely hand to a machine and do it unaided, once, this week. Write the difficult message yourself. Work out the number yourself. Plan the route in your head before you look. Notice two things: how much harder it is than you expected, and how much of that difficulty is the skill being rusty rather than the task being hard. That gap is the measurement. It’s the only one you’ll get, and it’s only available while you can still remember what the skill felt like.

People keep asking whether we could turn these systems off, as though the answer were technical. It is not. The switch works fine. What matters is the invoice attached to pressing it, and that invoice has a line item added every quarter by people who are just doing their jobs well.

An earlier post in this series looked at why the industry itself can’t stop building. This is the other half, and it’s the half that applies to you rather than to a lab. Even if every company stopped training tomorrow, the systems already deployed are woven into operations that a modern country runs on, and the weaving is the point. Nobody deploys a system into something unimportant. You deploy it where it saves the most, which is always where the volume is, which is always something that matters.

Follow the logic in an ordinary sector. A hospital group adopts a scheduling and triage system. It’s better than the old process, so waiting times fall and the group reorganizes around the improvement: fewer coordinators, tighter staffing, beds planned to the new efficiency. Two years in, switching it off doesn’t return the hospital to how it was two years ago. It returns the hospital to how it was two years ago while carrying today’s patient load with today’s staff numbers, which is not a return. It’s a crisis. The old system wasn’t kept warm, the people who ran it moved on, and the slack that used to absorb the difference was spent as savings, because leaving slack unspent is what a badly run organization does.

That’s the general shape and it applies everywhere. Efficiency gains get consumed, not banked. Every improvement is immediately spent on doing more with less, which is exactly what improvement is for, and the consequence is that the improvement stops being optional the moment it’s absorbed. Dependence isn’t created by the adoption. It’s created by the reorganization that follows.

Then multiply it by the fact that there is no switch. There are ten thousand switches, in different buildings, owned by different companies in different countries, each one controlled by someone whose job depends on their part working. Ask them to shut down and every one of them will point out, correctly, that their piece isn’t the dangerous one, that their competitors won’t stop, and that stopping would hurt their customers immediately and for certain. The coordination problem this series described for the labs reappears in miniature, in every organization, all the way down. There has never been a lever that turns off a general purpose technology, and looking for one is a category error we keep making because a single switch is a comforting object to imagine.

What worries me isn’t that the option gets taken away. It’s that it expires quietly while remaining formally available. You can always shut it down, in the sense that a person on a bridge can always jump. The capability stays real and the price rises until pressing it becomes the kind of thing only a catastrophe could justify, and by then the catastrophe has to be visible and enormous, and this whole series has been about why the thing that would justify it might not look like anything at all.

Fairness. Some of this is normal and fine. We couldn’t turn off electricity or the phone network either, and nobody loses sleep over it, because those systems don’t do anything except what we ask. The relevant difference isn’t the depth of dependence, it’s what you’re depending on: something inert and understood, or something grown, unreadable, and improving on a schedule nobody controls. And there are real mitigations, which deserve naming rather than dismissing: keeping manual fallbacks exercised rather than documented, deliberately preserving slack, having more than one supplier, running the drills. Aviation and finance both do versions of this and it works. It’s unglamorous and it’s the actual answer, and almost nobody is paying for it because the payoff is a disaster that doesn’t happen.

Tonight’s exercise, and make it concrete. Pick one system your work depends on, ideally one that arrived in the last three years. Now plan a two-week outage. Not the technical failover, the actual work: who does what by hand, whether those people still exist, how much of the volume you’d simply drop, what you’d tell customers. Get to a real plan or get to the honest sentence, which is that there isn’t one. Then note the date, because next year the same exercise gets harder, and the year after that it stops being an exercise.

It’s 2029, and Ruben approves about eleven hundred routing decisions a shift. Cargo, mostly. Which container goes on which vessel, which load gets bumped when a port slows down, which customer waits. His title is Operations Controller and his badge opens a door on the fourth floor, and he is, on paper, the person who decides.

His first year, he read them. All of them, at first, then most, then the ones the system marked as unusual. He caught two real errors that year. One was a duplicate booking and the other was a temperature-sensitive load routed through a transfer that would have taken nine hours in the sun. He still tells that story at dinners. It’s a good story, and it’s six years old.

The volume went from three hundred a shift to eleven hundred over four years, and the error rate went the other way. He hasn’t found anything in nineteen months. Not because he stopped looking, exactly. Because there stopped being things to find, and a person can only hold a posture of suspicion for so long against a thing that’s right every time. Vigilance isn’t a virtue you can decide to have. It’s a resource, and his ran out the way water runs out, gradually and then not at all.

There’s a dashboard. Of course there’s a dashboard. It shows average review time per controller, and Ruben’s is 1.8 seconds, and he is aware, without ever having been told, that 1.8 is good and that the woman two desks down who runs at 4.1 has been asked about it twice. Nobody has ever instructed him to be fast. The number simply exists, in a color, on a screen his manager looks at on Mondays.

He held one back last spring. Something about a consolidation looked off, and he flagged it, and it went to review, and review took eleven hours, and the shipment missed its window, and the client complained, and the system had been right. Nobody was unkind about it. His manager said good instinct, wrong call, that’s the job. Ruben agreed and went back to work, and he has not flagged anything since, and if you asked him whether that experience changed his behavior he would tell you honestly that it didn’t. He’d believe it, too.

On a Tuesday in October, decision 4,417 of his shift is unusual. A chemical load takes a routing through an inland hub that hasn’t handled that class before, on paper for a cost reason that’s real, under a permit that’s technically valid. It’s the kind of thing that would have made twenty-five-year-old Ruben pick up a phone. He approves it in 1.7 seconds, along with the four hundred before it and the six hundred after.

Nothing catches fire. That’s important, and it’s the part people get wrong about this kind of story. There’s a delay, a fine, and an insurance conversation, and the total cost lands somewhere between annoying and forgettable. An incident report gets written the following month, and it’s thorough, and it identifies the process gap, and it contains one sentence that Ruben reads four times before closing the file.

Human approval was obtained.

That’s the sentence. It’s true. It cost 1.7 seconds and it satisfied every governance requirement the company has, and it will appear in the audit as a control that functioned. And sitting at his desk at the end of the shift, Ruben works out what his job has actually been for the last three years, which is not to catch errors, because he doesn’t, and not to add judgment, because there’s no time for judgment at 1.7 seconds. His job is to be a name. The process requires a human to have approved, so a human approves, eleven hundred times a shift, and the requirement is met, and the requirement was never really about him.

There is no Ruben, and I made up the containers and the fourth floor. I didn’t make up the rest. Approval interfaces with volume beyond any possibility of review exist right now, in banks and hospitals and warehouses. Dashboards that time the reviewer exist. The rule that a decision must have human sign-off exists in a great deal of regulation, and it is satisfied, today, by clicks that take less time than reading the first line. The phrase human approval was obtained is doing enormous work in a lot of places, and almost nowhere does anyone check what it means.

So tonight’s exercise. Find the place in your own organization where a person is required to approve something at a volume no person could actually review. Every organization has one. Now ask what that requirement is protecting: the outcome, or the paperwork after the outcome. Then ask which one it was designed to protect.

No organization has ever sent a memo announcing that the machine now decides. What happens instead is a sequence of four small procedural changes, none of which is worth objecting to, and at the end of them the machine decides.

Step one, the system offers an opinion. It scores the loan application, flags the scan, ranks the candidates, and a person looks at the score alongside everything else. Nobody’s authority has moved. It’s an extra input, and extra inputs are free.

Step two, the opinion becomes the starting point. The form arrives pre-filled with the recommended answer, because pre-filling saves time and the recommendation is usually right. Notice what shifted. Before, you produced a decision. Now you edit one. Every study of defaults ever run says the same thing about what happens next, and it isn’t that people carefully evaluate the default.

Step three, disagreement requires a reason. There’s now a box: if you’re overriding the recommendation, briefly explain why. This is introduced as good governance, and honestly it sounds like good governance. But the burden of proof has just moved across the table, and that’s the whole ballgame. Agreeing costs nothing. Disagreeing costs a paragraph, a record, and a small implication that you might be the problem.

Step four, the override rate becomes a metric. Someone notices that overrides are correlated with worse outcomes, which they will be, because the machine is usually right and the humans overriding it are a mix of good instincts and bad days. So the override rate gets reported. Then it gets compared between teams. Then a manager mentions to somebody that their numbers are unusual. Nobody ever forbids disagreement. Disagreement just becomes a thing that shows up on your review.

At the end of those four steps, ask who decides. Legally and formally, the human does, and every process document says so. Practically, the human agrees, at a rate approaching one, for reasons that are individually sensible at every single instance. The authority didn’t transfer. It eroded, which is a different verb and leaves no paperwork.

The domains where this is running are not obscure. Credit, insurance pricing, medical triage, hiring, benefits eligibility, tenant screening, fraud flags that freeze your money, and in more places than most people realize, decisions inside criminal justice. What those all have in common is high volume, real consequences for one person at a time, and a decider who is overloaded and would genuinely benefit from help. That’s not a coincidence either. Decision support goes exactly where humans are drowning, which is where it’s most useful and where it’s least likely to be questioned.

I want to be careful not to write the lazy version of this argument. The lazy version says the systems are biased and the humans were fair. Often it’s the reverse. Human decisions in these same domains have been arbitrary, inconsistent, and worse before lunch than after it, and a system that applies one standard to everybody can be a genuine improvement for the people on the receiving end. There’s a real case that a well-built system beats a tired person, and in some settings I’d take the machine myself.

My worry is narrower and survives all of that. It’s that we get the improvement and lose the accountability at the same time, and only one of those was on the invoice. When a person decided, there was somebody who could be asked why, who could be argued with, and who could be wrong in a way that produced a consequence. When a recommendation is rubber stamped, the question why lands on nobody. The human says the system flagged it. The vendor says the human decided. The system says nothing, because it isn’t the kind of thing that answers. Every appeals process ever built assumes a decider exists, and the four steps above quietly dissolve one without ever removing anybody from the org chart.

Tonight’s exercise, and it’s a spotting game rather than a worry. Find a place in your own working life where you’d need to write a justification to go against a system’s suggestion. Not forbidden, just costly. It might be a scoring tool, a routing rule, a compliance flag, a forecast. Now ask yourself, honestly, how many times in the last year you paid that cost. If the answer is zero, that’s not evidence the system was always right. It’s evidence you don’t know whether it was, and neither does anyone else. Tomorrow, a short story about a man whose job is to click approve.

Ask any expert how they got good and you’ll hear a version of the same story: years of doing the boring parts, badly, while someone more experienced fixed them. Nobody has ever found a shortcut. Now notice that the boring parts are exactly what we just automated.

Expertise looks like knowledge from outside and it isn’t, quite. It’s calibrated judgment, and judgment is built out of a large number of encounters with the specific ways things go wrong. The senior lawyer who glances at a clause and says no is not recalling a rule. She’s recognizing a pattern she met eleven times as a junior, three of which went badly. The engineer who says that’ll break in production has watched it break in production. You cannot install that. It accumulates, slowly, through contact with real cases, and most of those cases have to be low stakes ones you’re allowed to get wrong.

That’s what the junior years are for. They look like cheap labor and they’re actually a training program disguised as cheap labor, subsidized by the firm because the firm needs seniors in ten years and there is no other way to manufacture one. The first draft nobody sends. The review of documents that turn out to matter twice a year. The model you build that your manager rebuilds. Dull, necessary, and the only known method.

So look at what these systems are good at right now. Not the hard judgment. The first draft. The document review. The starter model. The clean, well-specified, high-volume, low-stakes work with plenty of examples, which is the same list, item for item, as the junior year. This isn’t a coincidence. Those tasks were given to juniors precisely because they were the most structured and forgiving parts of the work, and structured and forgiving is exactly what a machine learns first.

Now put yourself in the chair of the person deciding next year’s hiring. A junior costs real money, produces little for eighteen months, needs a senior’s time to supervise, and might leave the moment they’re useful. The alternative is cheaper, faster, available immediately, and doesn’t need mentoring. There is no version of that comparison where hiring the junior wins on this year’s numbers. The manager isn’t being callous. They’re being correct, within the boundary of what they’re accountable for.

And that’s the trap, because the thing being consumed isn’t owned by anyone. A firm that skips its juniors gets a better margin now and a shortage in a decade, by which time it can hire experienced people from somewhere else. Every firm reasons the same way, and there is no somewhere else, because somewhere else was doing the same arithmetic. The supply of expertise is a commons that every individual actor is rewarded for not maintaining, and there’s no committee anywhere whose job is to notice.

What worries me most isn’t the unemployment. It’s the supervision. This whole series has leaned on the idea that humans stay in the loop and check the machine’s work, and checking requires someone who could have done the work. A reviewer without judgment isn’t a check, they’re a signature. So if the machines take the rungs, the next generation arrives at the review chair having never developed the thing the chair exists to apply, and the oversight that everyone’s safety plan depends on becomes ceremonial about fifteen years before anyone notices it happened.

The fair counterargument is that training changes when tools change, and it always has. Nobody learns navigation the way a ship’s officer did in 1850, and modern officers are not worse at their jobs. New tools bring new apprenticeships: maybe the junior year becomes about directing systems, catching their failure modes, and building judgment at a higher level of abstraction. That’s plausible. It’s also currently a hope with no curriculum attached, and I notice that the firms most enthusiastic about it are not the ones opening training programs.

There’s a second and more uncomfortable possibility, which is that judgment at the higher level of abstraction can’t be built without the lower one, the way you can’t develop taste in architecture without ever having stood in a building. If that’s true, the new apprenticeship doesn’t exist and can’t be designed, and we find out in about a decade.

Tonight’s exercise. Think back to how you actually became competent at whatever you do. Find the specific dull task you did hundreds of times, the one that felt like a waste at the time, and be honest about what it taught you that nothing else did. Then ask whether a person joining your field this year will get to do it. If the answer is no, the follow-up question is the whole post: who is going to check the machine’s work in 2040, and where exactly are they learning how.