November 05, 2025 • work, economics, automation

I have been putting off writing this for a month, mostly because I did everything I was supposed to do and it still feels like I have no idea what game I am actually playing.

I went to university. I got good grades. I did three internships. I ran a tiny consultancy for a while, building things for people and sending invoices that were actually paid. I studied computer science. I have done the very normal, very legible technical work you put on a CV: I learned the tools I was told to learn. I watched the right talks. I followed the right people. I can point at a neat little row of experiences and say: I played by the rules you told me about.

I am unemployed anyway.

I have been applying around as a new grad, sending out the same careful batch of CVs and cover letters that older friends used a few years ago to land solid jobs. What I meet instead is a job market that people now casually call broken. The phrase “white collar recession” has escaped into the mainstream. Articles talk about a graduate jobpocalypse and the disappearance of entry level roles just as the largest cohorts of students emerge from university. Computer science, which was marketed for years as the safest degree in the room, now shows up in headlines as one of the majors with the highest unemployment. People in my cohort are not confused about this the group chat tone is not a bit of a rough patch but this market is cooked.

On the official dashboards unemployment is still low, which is what older people tend to quote back at you. From the ground the thing feels different. The postings are there, the interview loops still exist, recruiters still send polite rejections. It is the density of opportunity that has changed. There are more people stacked against fewer real openings, and the default advice of “just apply to more places” lands differently when you know you are running through the same funnel as thousands of other people who also did everything right.

When you talk to older people, including many who are sympathetic, they often reach for familiar explanations first. Rates went up. Funding dried out. The last decade of cheap money was a sugar rush. The current US administration is running a particular mix of trade and industrial policy. These are all real. High interest rates do kill hiring sprees. Entire departments exist today mostly because capital was cheap three years ago. But behind that cycle there is something else, less visible in the graphs and much more visible if you are twenty something and trying to get your first proper job at exactly the moment capital has discovered that software and robots and offshore labour can be stacked together.

A decade ago people passed around a famous paper that said roughly half of American jobs were at high risk of computerisation over the following couple of decades. It was quoted so often that it turned into a kind of folk memory; a big, scary claim that sat behind every newspaper headline about robots taking work. The follow up work was much less memorable. When the OECD re-did the exercise at the level of tasks rather than whole occupations, the share of jobs flagged as “high risk” shrank to a much smaller slice, and later work showed that employment in those supposedly fragile occupations still grew, just more slowly than elsewhere. Automation as actually measured looked more like a slow pressure on certain kinds of work and wages than a cliff edge.

That slow pressure matters. Industrial robots in the United States have already been associated with sizeable job losses and clear wage falls in the regions where they were deployed. The OECD data shows that occupations with more routine, codifiable tasks have grown less and paid worse than occupations that rely more on social skill or physical presence. This is a story about composition, about how many of the jobs created are in fields people can actually enter and live on. The headline unemployment rate can stay low while a particular cohort feels like they are pushing against a wall that gets a little harder every year.

If you sit where I sit, as a new grad with a half decent CV and a browser full of ghosted applications, that gradual statistical story does not quite capture the feeling. What it feels like is that the entry corridor has narrowed, and that the bar for being worth a salary at all is moving faster than people want to admit. It feels like you are competing not just with other humans, which would be fine, but with the entire past of the economy: every dataset, every process flow someone has written down, every recording of someone doing the work you are trying to get paid for.

The Amazon story is a useful place to stare at this without drifting into science fiction. Leaked internal planning documents and external analyst notes have sketched out a future in which Amazon replaces a large share of warehouse tasks with robotics over the next decade and saves eye watering sums in the process. The company disputes some of the job loss projections, points to new roles created, and likes to say that robots help humans rather than replace them. It is probably sincere and even correct in some places. It is also true that Amazon has quietly expanded its robot fleet to a huge number of units while overall headcount in the most automated centres has flattened or fallen.

What interests me is not just the robots, but the quiet rewrite of the rule that sits under them. For most of the industrial era, you could assume that any large physical operation, like a warehouse, would need a certain number of human bodies to move boxes and drive forklifts. Human labour was a kind of fixed ingredient. You might squeeze wage rates or offshore some tasks, but you did not start your business model by asking whether you could avoid hiring people at all. Amazon and other firms like it are now encouraged by shareholders and banks to ask that question first. Not how do we run this warehouse with people, but how many people can we get away with and where can we put them so that they add marginal value on top of systems that coast on software and steel.

Teleoperation makes this even stranger. A surprising amount of so called automation today is really labour that has been routed through a screen. There are Filipino workers sitting in offices in Manila wearing VR headsets, remotely steering shelf stocking robots through Japanese convenience stores. There are people in one country sitting at desks, driving forklifts in another country using multi screen setups and a steering wheel, stepping in only when the semi autonomous software gets confused. Security robots patrol office corridors with a remote human ready to take over through a tablet whenever anything looks off.

It feels like immigration without immigrants. The rich country gets the labour it wants at a wage that looks more like Manila than Tokyo, but nobody has to build new housing, merge school systems, negotiate over culture or passports. On the rich side, it can be sold as productivity, which plays well politically. On the worker side, it is another rung in the long ladder that runs from call centres to business process outsourcing to micro task platforms. The worker is still human, still fallible, still earning just enough to keep going, but geographically they are treated more like part of the network than part of the town.

There is another twist, which is that these teleoperated jobs rarely stand still. A lot of them exist not just to get the work done, but to collect data so that the work can later be done without the human at all. Humanoid projects make this explicit. Neo, a household robot that went viral this year, spends a lot of its actual working time in an “expert mode” where a remote operator pilots it through chores, opening doors and picking up objects. The company then uses those sessions to train its own control model, using both successes and failures as data. Tesla’s Optimus is being taught in a similar way; workers wear rigs and repeatedly grasp cups or wipe tables so that the resulting recordings can be used as samples for the robot to imitate.

This is familiar if you watched what happened with data work for self driving cars and large language models. Scale AI started around 2016 with hordes of people labelling images and LiDAR frames for autonomous vehicle companies. Within a few years those labels fed into perception systems and then into foundation models for driving, and the work that was once scattered across many companies as internal grunt work was mostly concentrated with a handful of providers. Teleoperation feels like the embodied version of that. The immediate job is to keep the warehouse or the road running. The secondary job is to produce training data so that the next generation of robots and models will need fewer of you. Ghost work in the physical world.

I am not the person in the VR rig or in the forklift chair. My world is the white collar side of this, the part where the work happens in code editors, notebooks, documents and meetings. The pattern rhymes. Over the past couple of years you can watch entry level roles thin out in tech, finance, consulting and similar fields that used to soak up computer science graduates. Reports talking about a white collar recession and about entry level jobs disappearing just as the largest cohorts of graduates hit the market. Computer science, once the safest bet in the room, now shows up in stories as one of the degrees with the worst employment outcomes, which would have sounded like a joke not long ago.

You can see the shape of it in small ways. Entry level job boards that used to be full of junior developer roles now skew toward mid level and senior. Graduate schemes quietly cut their intake and companies push harder on automation and AI tooling instead. Employers tell journalists that they are holding back on junior hiring and prefer to lean on experienced staff plus AI tools, or simply automate parts of the work that would once have gone to juniors so they do not need to open the req at all. The ladder is still there, but it has lost a few rungs, and the remaining rungs sit above a seething pile of people who all did what they were told.

There is a tight relationship here with the way humans and software scale. Humans have always had some narrow horizontal scaling. A good operator could, in most cases, cover more than one vehicle if they were in the right geometry. Supervisors can coordinate teams. A single person with good tools can now keep an eye on several semi autonomous robots, or fleets of trucks, or a room full of humanoids, intervening only when something weird happens. Software takes this logic and stretches it until it breaks the category. Once you have a strong model, you can copy it into as many agents as you can afford to run. Read the current run of agent papers and demos and you see systems built from many copies of the same underlying model, set up to argue with each other, negotiate, plan and execute as little societies. In that world the baseline worker is not a person and not even a single bot but a swarm of cloned minds sharing memory.

Managers are starting to adjust their habits around that. There are already public memos from large companies where leaders tell their staff that any request for headcount has to come with a justification for why an AI system cannot do the job. Shopify’s chief executive talked this way when he told teams to try AI first before asking for more people. Some firms that used to keep small armies of contractors now advertise themselves as “AI first” and quietly shrink the human pool while their products shift more tasks onto models. It is not that nobody gets hired at all. Certain roles are still snapped up. It is that the default answer has flipped. The question is no longer whether a model can cover the job that was going to exist anyway. The question is whether a human can justify their presence next to a stack of models.

This is where I keep coming back to a phrase that has been rattling around my brain for the past month: out of distribution humans.

Most work lives in the fat middle of a bell curve. Tasks repeat with small variations. Most graduate schemes are built around that fact. You take reasonably bright people, give them a handbook and a mentor, and let them climb a well mapped gradient. Shared service centres, call centres, warehouses, junior consulting rotations, entry level software roles, even a lot of legal and accounting work, all sit in that comfortable hunk of the curve where yesterday’s data is a very good guide to tomorrow’s tasks.

Models feast on that part of the curve. That is what they are trained on: logs, emails, historical cases, recordings of someone else doing the job, code repositories, scanned documents. If your work looks a lot like a large pile of past episodes, it is a short hop from playing them back to imitating them. The central question for future labour markets is not whether you are clever or diligent in some absolute sense. It is whether what you do is ordinary enough for a model to learn or strange enough to fall through the gaps.

An out of distribution human, in my head, is someone whose job sits far enough in the tail of that curve that it does not currently compress into training data. Maybe they work with genuinely novel problems. Maybe they operate at small scales or in messy physical situations where we do not yet have enough sensors. Maybe they have taste that is not easily reduced to click logs. They are not safe; nothing is. They are simply late on the automation curve. The system needs them until it can watch them for long enough and in enough detail that it can flatten what they do into data.

The obvious problem is that most people, including most conscientious, capable new grads, are not doing anything like that. Most of us are trying to get into the middle of the curve, into the part of the labour market that has historically been considered sensible and respectable. My three internships were not wild experiments. They were exactly the kinds of things you do when you are aiming for a normal job: some engineering, some product, some research. The consultancy I ran was tiny and real, but it lived on predictable work. That was the point I was building a CV that sat neatly in the centre of the distribution, because that was where the jobs were.

Now it feels like the centre is being hollowed out. Employers are still very happy to talk about skills and effort, but the quiet question under everything is: is your contribution weird enough that we cannot paste it together from a few agents and a cheaper worker in another country. If the answer is no, then even when you do get hired you have to live with the knowledge that your day job is essentially a labelling job. You are adding examples to the pile that will train your future replacement. Out of distribution humans are the ones who manage to stay just far enough ahead of the pile for long enough to have a career.

The political reaction to this has not caught up. The industrial nations of the twentieth century were built around the idea that work was the organising principle of life. Catholic social teaching talked about the dignity of labour. Socialist movements sang about the worker as a hero. Protestant infused capitalism turned productivity into a route to salvation. Even the centrist stripe of postwar politics treated a job as the main vehicle through which adults were meant to find status, income and a place in the world. This hung around through the neoliberal years, even as manufacturing shrank and services expanded. You can hear it every time someone from any mainstream party talks about “hard working families”.

The result is that a lot of our institutions still act as if giving everyone a job is the primary goal, long after the underlying economic logic has started to drift. You can see it in the way some regions subsidise employment programmes that barely produce anything useful, or insist that people physically come in to do tasks that could be handled in far leaner ways. You can see it in the zombie jobs that exist mostly so that a local unemployment statistic looks less embarrassing. You can see it in small things like states that keep petrol pump attendants around in an age where self service is trivial, just to keep people on the payroll and maintain a social script about service. These are not vast conspiracies but instead residual behaviour of a world that treated labour as sacred.

Unions sit in the middle of this. They have in some cases slowed automation in ways that probably preserved wages and bargaining power for longer than markets would have allowed. Metro lines in Europe run with drivers even though driverless lines exist in the same city and have been proven technically viable. Port workers have fought hard to restrict automated cranes and remote controls, sometimes winning explicit clauses that bar certain kinds of automation for the life of a contract. Strikes that shut down whole systems have been used as leverage to negotiate over the future of jobs that could in principle be done with much less human involvement.

There is a strange symmetry here. On one side you have firms quietly routing labour through screens and robots, and repeating that jobs will be fine on aggregate. On the other you have unions and politicians insisting that jobs must be preserved, even when that means attaching people to tasks that are technically obsolete. Neither camp really articulates what it would mean for work itself to shrink as a central organising story. They just fight over where the remaining jobs will be and who will do them.

If you want to see how far this can go without the sky falling, you look at places that have already pushed automation hard. The International Federation of Robotics statistic tables are slightly dry, but they tell a simple story. South Korea, Singapore, Japan and Germany have been packing industrial robots into their factories at a remarkable rate for years. China started later, then began installing new units at such a pace that it now accounts for more than half of global industrial robot installations and has overtaken Germany in robot density in manufacturing. At the same time, China’s GDP per capita by purchasing power parity is still maybe a third of that of the United States, and its youth unemployment has lurched upwards, with official figures in recent years hovering in the mid to high teens and unofficial estimates higher.

So you have a giant country that has thrown money and policy at automation, packed factories full of robots, tied itself into global supply chains, and still produced a generation of graduates who complain on social media about “rotting” in low paid service jobs or online hustle work. You have very visible memes about lying flat and giving up on the competition game. You have an official narrative about hi tech growth that is not exactly false, but feels distant from the daily experience of a twenty three year old with a degree and no good offers.

The gig economy in the United States and Europe is another small window into this. Robotaxis are still barely a rounding error in total miles driven. Waymo carries a trivial share of all rides in the cities where it operates. Yet if you talk to ride hail drivers in San Francisco or Phoenix, you hear anxiety that grows faster than the fleet. Data from driver apps already shows earnings slipping in markets where robotaxis operate even though their absolute presence is small, and banks have issued notes warning that urban ride hail platforms face “AV risk” as robotaxi coverage expands. The job loss story gets there before the job loss itself.

This is the pattern that worries me more than the big forecast numbers. The technical line is that automation is a slow, uneven pressure; that jobs appear as well as disappear; that productivity can even raise wages in the long run. The lived line for my cohort is that the good jobs at the centre of the curve are thinning out, that junior entry points are being quietly closed, and that we are being told to somehow mutate into outliers while competing with systems that learn from everything we do.

I do not know how many jobs will exist in twenty years, or whether my own work will sit far enough into the tail of the distribution to matter. I will certainly try to become an out of distribution human by doing a lot of different things, and by refusing to live entirely in the centre of the curve but if your entire life plan rests on being a respectable, central case worker, doing a standard job in a standard company, I think you should at least stare straight at how much effort is going into eroding that category. If your politics rest on the idea that everyone will work full time and find dignity there, you should stare at it too. The twentieth century spent a lot of intellectual and moral effort glorifying labour because economies needed people to show up every day. The twenty-first century is starting to build machines and systems that do not need quite as many of us.