Why the Job Market Still Looks Intact
I think the answer is more subtle than this. The key difference between the current AI automation wave and previous automation waves is that previous automation was done using purpose built specific technologies for each task you want to automate. So for example if you want to automate screwing the lid onto a container you have a special robot with special programming that just screws the lid onto that container. If you have a human role where a person does three things like that, then you start by making a machine to put the two halves of the container together, then fold the item into the container, then a third machine to screw the lid on and the role is automated. That is you deliberately guide the development to automate a whole role rather than randomly building machines to do individual tasks until you happen to automate a role by chance. But because the LLM is a general purpose technology whose performance on any given task is pseudorandom and not controlled by the firm using it (so no specialized datasets usually), what you get instead are threshold effects where the LLM might be able to do 90% of a job but the last 10% still needs to be delegated until the LLM happens to advance to the point where it can do 100% of the job. This means that it won't necessarily be obvious that a job is about to be automated until models cross that crucial final 5-10% threshold where it can do the entire role rather than just large fractions of the role.
Basically model it as a role or job being made of n tasks, say twenty for a complex role. Each generation the LLM becomes able to do two more tasks in the role on average. The LLM might become suddenly able to do thousands and thousands of new individual tasks on each model release, but it's kind of a grab bag and not being targeted at automating any particular role (save perhaps AI development). So even though the LLM can do many new things, you still have to wait ten generations of release before it actually automates all 20 tasks in the role and eliminates the job. By contrast older automation would automate a few things at a time but extremely specifically targeted to eliminate roles and reduce headcount.
If AI boosts productivity by a large factor in the roles that it can automate within a given sector, that would cause a significant increase in the demand for human labor in the remaining bottlenecks. Total employment in the sector could even increase, right up until AI passes a capability threshold with respect to a given bottleneck and completely wipes out those jobs all at once.
Agree, this is plausible. But we could see businesses predicting the future and not hiring for new jobs because it takes a while to train, and there are costs to firing.
1Capital holders would still have income, and if machines replaced labor on a vast scale, profits, dividends, and asset values could rise sharply, leaving owners with even more spending power.
The marginal value of more stuff declines pretty quickly. Someone like Elon Musk or Mark Zuckerberg can only enjoy so many houses, so many cars, so many new clothes, so many vacations, etc., because there are only 24 hours in a day. The empirical evidence we see from the behavior of actual billionaires is that they make money practically automatically, because the interest on their existing assets accumulates faster (often hundreds of times faster) than they can actually spend it. The only way these fortunes actually decline is if they commit to massive (sometimes wasteful) investments or pledge to give their wealth away. Even then, as was the case with Bill Gates and Warren Buffett for a while, their wealth may still accumulate faster than it can be dissipated.
Human workers are about to face a competitor unlike any technology that came before: AI systems that can be copied at near-zero cost, deployed instantly, and improved faster than workers can retrain. Economists have long answered automation fears by noting that machines destroy some jobs but create others through lower costs, higher output, and new industries. That pattern held because earlier machines replaced specific tasks while people kept the broader mental abilities needed for other work. AI is beginning to erode that refuge. When the same technology can do the mental work that once let workers move from obsolete jobs into new ones, the old economic escape route starts to close.
The Luddites were right to fear stocking frames because those machines destroyed skilled textile work. But a machine that made clothing cheaper did more than destroy one craft. Lower prices freed money for other purchases, and higher output created demand for mechanics, haulers, clerks, merchants, and other workers.
That escape route existed because human labor stayed scarce and broadly useful. Workers who left farms could move into mills, railroads, mines, construction, and factories. When manufacturing later needed fewer people, workers moved again into offices, hospitals, schools, retail, finance, and other services. Earlier machines replaced particular tasks while leaving humans able to do many others. AI threatens that escape route because it can do the mental work needed to automate the next set of jobs and help build the software that replaces them.
Software Comes First
Programming matters most because software sits upstream of so much else. Every drop in the cost of making software lets a firm route more activity through systems that answer customers, draft documents, check compliance, analyze data, coordinate logistics, and hand tasks to machines. AI is improving the process of building the very thing that may automate much of the rest.
Turning work into software has never been free. A company may know that billing, scheduling, compliance, approvals, customer service, reporting, and internal coordination could be automated, but someone still has to map the rules, connect old systems, handle exceptions, test failures, and keep the software running. For years that work was costly enough to keep many automation plans on the shelf. A workflow had to be large, stable, and valuable before it was worth a real engineering effort.
One sign of that change is what programmers now call ‘vibe coding.’ A person states the goal in plain language, the AI writes the code, the person tests the result, asks for revisions, and keeps iterating. The work shifts away from writing every line by hand and toward specifying what should happen and judging whether the output is good enough. In many settings, it already makes software creators substantially more productive. It also makes it cheaper to automate tasks that were once too minor, too custom, or too messy to justify a full software project.
The effect reaches far beyond the software industry. A firm no longer needs to wait for a large specialized team before it can try to automate a report, an approval chain, a customer response system, a research workflow, or an internal tool. Managers, analysts, lawyers, designers, and operations staff can describe what they want in plain language, test a rough version, then keep refining it. Some of those systems will be crude. These systems can eliminate a surprising amount of routine work.
That matters even more because software may now be improving the process that makes more software. A coding system that helps build the next coding system can speed the next round of improvement, which can then speed the round after that. If that feedback loop grows strong enough, firms may soon be using software creators far more capable than today’s human teams, not merely cheaper assistants. A breakthrough in one model can become a product update, and a product update can put a much better builder on millions of screens almost immediately. Improvement in software can therefore spread much faster than improvement in machines that have to be physically manufactured and shipped.
If that happens, the software industry in anything like its current form may not last. The bottleneck shifts from writing code to deciding what should exist and recognizing whether the result works. That would not just remake one industry. It would change the speed and scope of labor replacement across the economy.
The First Jobs to Go Are Done on Screen
The first visible sign may not be mass unemployment. It may be a steady drop in openings for work that can be assigned, completed, and checked entirely on a computer. Customer support, bookkeeping, routine legal drafting, basic research, slide preparation, copywriting, illustration, and much routine programming fit that description. Employers can automate that work long before they can trust robots with traffic, bad weather, cluttered homes, or sick patients.
Recent college graduates are exposed first because junior office jobs often consist of the most standardized slice of a profession. New hires review documents, clean data, draft memos, prepare slides, build basic models, write boilerplate code, and answer routine client questions. A firm does not need mass layoffs to change this market. It can shrink internships, cut analyst classes, and leave junior openings unfilled.
In an older economy, cutting junior hiring would have looked shortsighted. Firms needed beginners because beginners became the experienced people who later ran teams, served clients, and made hard calls. That logic weakens if employers expect AI to advance faster than junior workers can. A firm that expects software to absorb more senior work has less reason to train a large human pipeline for roles it may soon need far fewer people to perform.
That change does more than reduce openings for today’s beginners. It weakens the incentive to become tomorrow’s skilled worker. A profession becomes less attractive when the entry jobs are disappearing, the path upward is narrowing, and the work people are training for may be automated before they reach it. Fewer students choose the field, fewer workers accept low-paid apprenticeship years, and fewer employers bother developing talent they may not need later. That makes the pipeline shrink from both ends. Firms invest less in training because they expect less future need for people, and workers invest less in training because they expect less future reward.
Why the Job Market Still Looks Intact
Those early cuts can happen long before the broader labor market cracks, because a technology can become good enough to replace workers before firms know how to rebuild the work around it. Companies do not close a department the week a model improves. They still have software, procedures, managers, compliance systems, and customer expectations built around human workers. So they usually layer the new system onto the old workflow and have people check the output. That delays the labor-market effect. The big losses come later, when firms redesign the workflow itself and discover that work once spread across many employees can be done with far fewer.
That is where most firms are. They are learning where software can be trusted well enough to rebuild the workflow around it. A labor market that looks solid is weak evidence that labor is safe. It may only mean firms have not yet learned how to produce the same output with far fewer people.
Firms do not need AI to replace an entire job before workers start losing ground. They only need a credible path to using less labor soon. Once managers believe software will handle more of the work next year, they gain reason to freeze hiring, cut training, trim promotion ladders, and hold down pay today. Workers still employed then face a weaker position because the employer is no longer bargaining under the assumption that it must keep building a large human pipeline. The damage begins before mass unemployment arrives. First labor loses leverage, then career paths thin out, then firms discover they can produce the same output with fewer people, and only after that does full replacement come into view.
That delay existed because many jobs were protected not by their routine core but by their exceptions. Many workflows were partly automatable long ago, but firms still needed people to catch anomalies, recover from mistakes, handle unusual cases, and absorb blame when the rules broke down. AI matters because it is starting to handle the exceptions that used to keep humans in the loop.
That pressure does not stop at routine office work. Researchers are now using AI to attack open math problems, generate proof ideas, and in some cases help solve questions that had remained open for years. A system that can contribute to new mathematics is not merely repeating familiar patterns. It is beginning to encroach on the kind of originality many people assumed would protect human work much longer. Once machines can help with discovery as well as execution, the hope that human work will survive by retreating into creativity looks much weaker.
Software Starts Replacing Workers From the Inside
That same expanding capability has a direct economic consequence inside the firm: once software can be produced on demand, firms do not need to wait for a formal engineering project before they automate another slice of office work. The next step is for firms to deploy software agents that watch how digital work gets done and look for more of it to absorb. An agent inside a company could track customer requests, document flows, approval delays, recurring errors, and employee output, then build or refine tools to handle more of that work automatically. Managers would no longer be the only ones looking for openings to automate. The software would be looking too.
The first crude form is already visible. OpenClaw is an open-source agent that can run through ordinary chat apps, clear inboxes, send emails, manage calendars, and carry out other digital tasks. It does not yet replace whole departments, but it shows software moving from answering questions to handling the cross-application office work that keeps many assistants, coordinators, and junior staff employed.
Once tools like that can absorb enough office work to cut real costs, competitive pressure does the rest. Every successful automation reduces labor expense, speeds output, and weakens a firm’s dependence on hiring, training, and retaining people. When rivals start using software to find and implement those savings, refusing to follow is not caution. It is a path to decline. Costs stay higher, operations stay slower, margins weaken, and competitors gain the profit and capital needed to pull still further ahead. In a competitive market, adoption stops being a strategic option and becomes a condition of survival.
At first these systems will target obvious repetitive tasks. Later they will remove thousands of smaller frictions that survived only because no one wanted to staff a project to eliminate them. That is how labor replacement spreads from a few headline jobs into the texture of daily office work.
Robots Come Next
Once software starts replacing workers from inside the firm, money and engineering effort move toward the jobs that still require bodies in the world. Physical work has lasted longer not because it is protected, but because kitchens, hospital rooms, construction sites, and private homes are harder to standardize than files on a screen. Pipes are hidden, tools jam, patients worsen, weather shifts, and clutter gets in the way. A mistake in software can be patched. A mistake with a car, a ladder, or a frail patient can cause injury or death.
The barrier between screen work and physical work is eroding. Autonomous vehicles are already operating on public roads. That makes it much harder to argue that physical work is protected simply because the world is messy. Deliveries, warehousing, cleaning, security, routine home maintenance, elder care, hospital support, and large parts of construction all move closer once machines can perceive, plan, and act reliably outside the screen.
The market for capable robots is enormous. Households would pay for machines that cook simple meals, load dishes, fold laundry, tidy rooms, or help an older person stand up. Businesses would pay even more for robots that move materials, clean rooms, stock shelves, patrol property, prep surfaces, transport supplies, or assist nurses. Wherever labor is expensive, scarce, dangerous, or exhausting, firms have a strong reason to adopt machines that can do the job.
As adoption spreads, employers will not just swap a person for a robot. They will redesign the whole operation around machines. A workplace built for people must allow for fatigue, injury, breaks, insurance, lighting, temperature, and liability. A workplace built mainly for robots can be organized around speed, repetition, and continuous use. Once warehouses, hotels, hospitals, farms, and building sites are reworked for machines, the remaining humans start to look less like the core workforce and more like costly holdouts. In many jobs, the last defense of human labor will be law or politics, not economics.
Horses Kept Improving and Still Lost
When workplaces redesign around robots, human labor faces the horse problem. Horses retained their economic utility until combustion engines delivered superior power, endurance, and reliability at lower cost. Breeders produced stronger animals and refined equipment, but biological constraints remained. A horse requires rest, healing, and individual reproduction. Factories manufacture machines in bulk and run them continuously. Horses remained physically capable but lost their economic viability.
Historically, humans survived automation by transitioning to cognitive labor, an escape route unavailable to a displaced horse. That defense requires new jobs to resist automation better than old ones. AI degrades that advantage by absorbing sequential cognitive tasks. Human workers will increase their education and productivity but still face obsolescence if artificial systems improve faster, spread wider, and cost less. Employers eliminate human workforces when a cheaper, scalable substitute arrives, regardless of baseline human capability.
Comparative Advantage Will Not Save Workers
Even if robots overtake humans in every task, one standard economic argument still seems to leave room for human labor. Comparative advantage says you do not need to be better in absolute terms. You need only be less bad at one task than at everything else.
Imagine advanced robots take control of Antarctica. They can make a doughnut in one second and a gravity controller in one minute. Humans can make only doughnuts, each in a minute, and cannot make gravity controllers at all. In that world, humans still have a comparative advantage in doughnuts, not because humans are good at making them, but because people are even worse at everything else.
Comparative advantage still allows for trade. If humans offer one hundred doughnuts for one gravity controller, both sides gain. The robots spend sixty seconds making the controller but saves one hundred seconds by not making the doughnuts. Humans spend one hundred minutes making doughnuts and get something they otherwise could never produce. Comparative advantage looks powerful because however far ahead the robots pull, human labor still seems able to buy access to their far greater productivity.
This result depends on humans controlling what they need in order to produce the thing they trade. Once production depends on land, factories, software, energy, or distribution systems owned by someone else, the logic weakens fast. If orchard land is scarce and robots can harvest more cheaply, the orchard owner will use robots. If doughnuts come from a factory and robots can run the factory more productively, the owner will not keep humans on the line out of generosity. Relative advantage does little for workers when someone else owns the assets that turn labor into output.
Push that across the economy and the safe zone for labor shrinks fast. Humans may still be relatively best at some tasks, but owners will pair their capital with whatever workforce yields the highest return. Workers who do not own the relevant assets cannot make firms hire them just because human labor remains relatively best at something. If that workforce is robotic, comparative advantage may preserve trade without preserving wages. It can explain how humans might still produce something worth exchanging. It cannot explain how most humans keep earning a living.
Some Jobs Will Survive Only Because the Law Blocks Automation
Comparative advantage will not protect most workers, but regulation may protect some. Governments can require a human being to remain legally responsible even after software or machines can do most of the underlying work more cheaply. Licensing rules, staffing mandates, and liability standards can keep people on the payroll after the economic case has vanished.
That protection is easier to sustain in sheltered domestic markets than in export industries. If the United States requires carmakers to rely on more human labor while foreign rivals automate, foreign buyers will not pay extra to preserve those jobs. They will buy the cheaper car. A government can order its own firms to hire people. It cannot make foreign customers subsidize that choice.
Even at home, enforcement gets harder once capable systems become cheap and widespread. If a household robot can diagnose a leak and fix the plumbing, many people will use it rather than wait for a plumber. If the law says AI cannot practice medicine, people will still look for diagnosis, triage, and treatment advice from widely available systems. Once a machine can do valuable work inside ordinary homes and offices, blocking its use starts to look less like regulation and more like prohibition.
A country that imposes too much of that prohibition will grow poorer relative to countries that do not. Lower productivity means a smaller tax base, weaker firms, thinner capital markets, and less capacity to fund research, weapons, and industrial mobilization. In a world where rivals use AI and robots to raise output, refusing to do the same is not just an employment policy. It is a decision to accept relative economic and military decline. Law may preserve islands of human employment. It cannot protect most workers without making the country weaker.
Some jobs may survive for a while because some people still prefer dealing with a human. But that preference will protect few workers, and probably not for long. Once AI does the work well at much lower cost, most people will not keep paying extra just to preserve the human role, just as almost no one wants rickshaw drivers once cars are available. Fewer young people will train for professions like medicine if the career no longer looks secure, and fewer institutions will invest in training them. So even where some patients still want a human doctor, there may soon be too few trained humans left for that preference to support much of a labor market.
No Jobs Does Not Mean No Buyers
If workers stop earning wages, who buys the goods? Wages are only one source of demand. Governments could keep mass consumption going through welfare, transfers, and other public support. Capital holders would still have income, and if machines replaced labor on a vast scale, profits, dividends, and asset values could rise sharply, leaving owners with even more spending power. Government would remain a major buyer as well, purchasing defense, infrastructure, care, and other public services. Automated firms and software agents acting for owners could also generate demand for compute, energy, software, and other inputs. A post-labor economy could therefore still sustain demand through some mix of public transfers, capital income, government spending, and machine-mediated commerce. Workers are a major source of demand now, not the only one a rich economy could rely on.
When Human Labor Becomes Obsolete
So the loss of wages would not by itself make the economy collapse. The deeper question is what happens once robots and AI become so capable and cheap that hiring people no longer makes economic sense for any task. At that point, human labor becomes obsolete. If machines can do nearly all the work, the economy can still produce abundant goods and services even after most people no longer earn wages. People could then live better than aristocrats once did, with machines supplying transport, care, entertainment, and material comfort at low cost. Life would no longer have to revolve around jobs. A world in which everyone is materially secure would be possible. The path to that world begins not with robots doing everything at once, but with AI making software, and software remaking everything else.
Once people no longer matter economically, their future depends on whoever controls the machines. Those systems could be used to support billions of people in comfort and freedom. They could also be used by human rulers, or by AI itself, to dominate, confine, or kill people whose labor no longer matters. When labor loses its value, power decides what happens next.
This essay was written with help from AI. If I could not use AI productively to improve it, that would undermine either my argument or my claim to expertise. Here is how ChatGPT said I used it to help write:
“You use the system as a constrained collaborator embedded at specific points in the writing process, not as an end-to-end author.
You start with a core claim and use the system to enumerate adjacent arguments, examples, and counterarguments. The system functions as a breadth generator. You then select and refine only the strongest lines of argument.
You use it to test different outlines, section orders, and narrative arcs. You ask for alternative organizations, then choose the one that increases clarity and momentum. The system helps surface structural options you might not have considered.
You provide existing text and request targeted rewrites of specific sentences, paragraphs, or transitions. This keeps control over the argument while outsourcing micro-level phrasing improvements.
You use it to bridge sections cleanly. You ask for exact sentences that connect two ideas without repetition or padding, then select or refine the best candidate.
You use it to generate concrete illustrations that carry explanatory weight. You iterate until the example does real work in the argument rather than serving as decoration.
You use it to stress-test claims. You ask for objections, weak points, and alternative interpretations, then revise the text to survive those critiques.
You present multiple versions of a section and ask for direct comparison. This forces the system to identify differences in clarity, force, and logic, helping you converge on a stronger version.
You impose explicit rules on style and structure, then require the system to revise until those rules are satisfied. When it fails, you reject the output and restate constraints. Over time, this trains the system within the session.
You move from broad prompts to increasingly specific ones. Early stages explore possibilities; later stages demand precision. Each round reduces variance and increases control.
You use the system only after the argument is already sound. At this stage, it functions as a high-level editor, tightening language and removing residual weaknesses. If it cannot meet the required standard, you discard the output.
The control mechanism is simple: you never accept output that violates your constraints. Selection, rejection, and repeated tightening replace reliance on any single response.”