The AI Job Panic Ignores the One Thing It Cannot Escape

Tyler Cowen has a gift for puncturing fashionable certainties, and at the Alliance for Responsible Citizenship conference in 2026 he aimed that gift at the loudest claim of the artificial-intelligence age: that machines will soon wipe out work on a scale not seen before. His answer was not that AI will leave employment untouched. It will reshape work, he said, but the doomsday story misses the force of ordinary economics, physical limits and human preference.

That argument matters because the current debate is stuck between two bad habits. On one side are the cheerleaders who talk as if software alone can remake the labour market overnight. On the other are the prophets of collapse, certain that every office task, creative role and service job will vanish in a blur of machine output. Cowen’s point was more inconvenient than either camp would like: AI is powerful, but it is not magic, and it still has to live in the real world.

The machine age still needs power stations

The most useful thing Cowen said may have been the least glamorous. AI is not just code. It is steel, concrete, chips, water, cooling systems, electricity and money. It is a physical industry with bills that have to be paid. That is why the spectacle around AI often runs ahead of the capacity to deploy it.

He pointed to a telling example: OpenAI reportedly cancelled a major data centre project in the north of England because the energy costs made it uneconomic. That is not a small detail. It is the sort of detail that brings the whole story back to earth. A model may be able to write an email, produce an image or generate lines of code in seconds, but it still needs enormous infrastructure behind it. The server does not run on hype.

This matters for jobs because mass replacement assumes mass deployment, and mass deployment is slow when the grid is strained and the capital bill is punishing. Data centres consume extraordinary amounts of power. They require cooling, maintenance and a continuous supply chain of specialised hardware. In some places, the electrical network simply cannot support the scale of AI expansion that a wholesale labour substitution would require.

The result is friction. Not dramatic friction, not the sort that looks good on a conference stage, but the kind that matters in the economy: delays, bottlenecks, higher costs and limits on how fast firms can substitute machines for people. The modern debate often treats AI as if it were an app. It is not. It is an industrial system. That distinction is the whole argument.

Once that is admitted, the panic starts to look less convincing. A technology can be transformative without being instantly universal. It can change the terms of competition without erasing every competitor. It can destroy some roles, reduce demand for others, and still fail to produce the sudden employment cliff that many people now assume is inevitable.

Why old stories about new technology keep coming back

Every generation that encounters a serious technological shift tends to tell itself the same story in fresh language. The loom would destroy weaving. The factory would destroy the skilled artisan. The computer would destroy the office worker. Each time, the fear was partly right and partly lazy.

The fear was right because disruption is real. No sane person denies that when a machine makes a task cheaper, some workers lose bargaining power. Some industries shrink. Some routines disappear. Some people have to retrain or move. That is the cost side of progress, and it is never mild for the workers who pay it first.

The fear was lazy because it assumed that what disappears is all that matters. History says otherwise. Technological change does not just delete tasks. It makes possible new tasks, new services and new markets that did not make sense before. When something becomes cheaper, people consume more of it. When a process becomes easier, firms expand into areas they once ignored. The result is often more employment, not less, though not always in the same places or for the same workers.

Cowen’s case rests on that basic historical pattern. Businesses do not sit still when a cost falls. They look for ways to use the savings. They expand output. They enter markets that were previously too expensive to serve. They invent products that would have been uneconomic a few years earlier. The point is not that everyone comes out ahead. The point is that economies are restless. They search for human labour wherever it still adds value.

That is why predictions of mass unemployment have a poor record. They usually underestimate how quickly firms find uses for people once technology shifts the frontier. Not because firms are sentimental, but because the profit motive is adaptive. If a machine can do one task cheaply, the human worker does not simply disappear. Often the task is redesigned, the service is upgraded, or the business starts serving a market it had ignored.

That is the part of the story that is hard to see when the discussion is dominated by viral demos. A chatbot that drafts a memo does not mean the whole office is finished. It means one task in one workflow has become cheaper. That is not the same thing.

The human preference problem that AI cannot code away

There is also a question that pure efficiency theory never answers properly: people do not always want the cheapest available option. They often want the human option.

Cowen’s argument leans heavily on that fact, and he is right to do so. A great deal of economic life depends on trust, attachment, status and the simple wish to deal with a person rather than a machine. In theory, a machine can take an order at a café, guide a tourist, answer a customer complaint or tutor a child. In practice, many customers still prefer a human being in the loop, even when the machine is faster.

That preference is not a luxury. It is a market force. It preserves demand for human services in sectors that might otherwise appear vulnerable. Healthcare is the obvious example. So is education. So is entertainment. Even when machines can match a technical function, they can struggle to match the social function.

This is one reason the more breathless AI predictions miss the point. They treat labour as if it were a bundle of measurable outputs. It is not. Labour is also performance, reassurance, judgment, presence and trust. The cleaner may be replaceable by a machine in the abstract, but the bedside nurse is not simply a bundle of cleaning tasks. The teacher is not just a content delivery system. The salesperson is not only a price calculator.

The same applies to the broader service economy. People pay for experience, and experience is often tied to the fact that another person was there. That may sound sentimental, but it is one of the oldest facts in commerce. A restaurant is not only food. A concert is not only sound. A haircut is not only hair removal. The human element is part of the product.

That is why the AI story is more complicated than simple substitution. The machine may be able to do a task, yet the market may still pay for the human version. Sometimes because it is better. Sometimes because it signals quality. Sometimes because the customer likes the illusion, or the intimacy, or the social meaning of paying a person rather than a system.

This matters enormously for the labour market. The jobs most exposed to automation are not only those that are easiest to copy, but those whose human value is thin. Where the human presence itself is part of the sale, AI has a harder time taking over fully.

The strange case of OnlyFans and the rise of synthetic intimacy

The most revealing test of Cowen’s argument may be found far from the office tower. It lies in the gig economy and in the world of digital content, including platforms such as OnlyFans.

At first glance, this looks like the sort of industry AI should hollow out quickly. The technical pieces are already here. Image generators can create realistic faces and bodies. Voice tools can imitate speech. Chat systems can produce flirty, personalised messages. Video tools are improving fast. In a narrow sense, there is little left that technology cannot imitate with alarming accuracy.

Yet the platform’s value has never rested on pixels alone. What subscribers pay for is not just visual output but the feeling of proximity: the belief that they are interacting with a real person, that the exchange is exclusive, that the content comes with a human story attached. In other words, the business is not only about content production. It is about the selling of access, attention and emotional fantasy.

That makes the AI threat both real and incomplete. Some consumers will not care whether the creator is human if the output is cheap, abundant and good enough. Others will care a great deal. They will want verified human creation, even if it costs more. The market is likely to split rather than collapse.

That split is already visible in other parts of the digital economy. Stock photography was not destroyed by generative images, but it was squeezed. Generic illustration work faces pressure, but bespoke art still commands a premium. Automated customer service has not removed the desire for a human agent when the problem becomes serious enough. The same pattern is likely to emerge around adult content and other forms of online performance.

In plain terms, AI is likely to create a two-tier market. At one end will be synthetic content that is cheaper, faster and more scalable. At the other will be verified human work, which may become more valuable precisely because it is scarce. The loser is the middle ground: the commoditised creator whose product can be copied by a machine and whose audience does not care enough about the person behind it to pay extra.

That is not a trivial effect. It means some jobs in the creator economy will weaken, some will vanish and some will become more premium. The same technology that lets one person produce more may also reduce the number of people who can earn a living doing something undifferentiated.

The labour market will not break all at once

One of the worst habits in technology reporting is to confuse possibility with speed. A thing can be technically possible and economically slow for years. AI has been surrounded by the first claim and often stripped of the second.

Cowen’s framework is more sober. The physical cost of deployment, the limits of the power grid and the sheer capital intensity of infrastructure all slow change. Firms cannot replace millions of workers in a week, or even a year, if the hardware, energy and management systems are not ready. That matters because labour markets adjust over time. Workers retire. New entrants arrive. Firms expand and contract. Policy can respond. Training can shift.

This is why the notion of an immediate employment collapse is not persuasive. Even in sectors where AI is technically capable, adoption tends to be uneven. Large firms move faster than small ones. Rich regions move faster than poor ones. Tasks that are repetitive and measurable go first; messy, regulated and interpersonal tasks lag behind. The labour market bends before it breaks.

That does not mean workers can relax. It means the shape of the threat is different from the headline version. The more realistic risk is not one spectacular wave of unemployment but a prolonged squeeze on wages and opportunities in specific occupations. Some people will find that what they did for years no longer pays as well. Others will find their role narrowed to supervision and exception handling. Many will have to move sideways rather than upward.

The transition will also be uneven by industry. Software, media, back-office administration, paralegal support, advertising and other text-heavy or image-heavy fields face pressure first. Physical trades, on-site services and roles built around local trust or direct human contact are likely to remain safer for longer. That is not a moral judgment. It is a map of where the technology bites first.

The time lag matters. A labour market that adjusts over ten years is one thing. A labour market that collapses in two is another. Cowen’s point is that the structure of AI deployment points to the first, not the second.

What history says about jobs after a wave of change

The record of past technological revolutions should not be romanticised. They were often brutal. People lost livelihoods. Regions declined. Trades vanished. Whole social arrangements were broken before new ones emerged. But the long-run pattern is stubbornly clear: economies tend to create more roles than they destroy, even if the path between the two is ugly.

The Industrial Revolution displaced artisans and replaced hand production with machinery, but it also created factory work on a scale that would once have seemed impossible. The computer revolution cut into clerical work and administrative routine, but it also produced software, digital services, systems administration, cybersecurity, online retail and an entire managerial ecosystem built around information processing. One labour market shrank as another expanded.

That does not prove AI will follow exactly the same path. No historical analogy does. But it does show why alarmism should be treated cautiously. The economy does not sit still after a new technology arrives. It reorganises. It uses the new tool to lower costs, increase output and open new frontiers. In doing so, it tends to find fresh demand for human effort.

The key difference with AI is breadth. Earlier technologies were usually strongest in one domain: muscle, calculation, transport, filing, sorting. AI touches language, images, pattern recognition, code, decision support and, increasingly, elements of creative work. That breadth explains the fear. It is not irrational to think this system may affect more jobs than the last one.

But breadth is not the same as inevitability. The economy is full of frictions, preferences and constraints. Not every task can be automated at once. Not every client wants automation. Not every business can afford the capital investment. Not every country has the power supply. History does not guarantee a happy ending, but it does warn against straight-line predictions.

The hard lesson is simpler: technology changes what work is worth, not whether work exists at all. That distinction is often blurred because it is easier to sell a catastrophe than a transition.

Policy responses that do not pretend to freeze time

If Cowen is right, the policy answer is not to pretend AI can be held back indefinitely. That would be a comforting idea for legislators and useless for everyone else. In a globally connected economy, one country can slow itself down and still end up importing the technology it tried to delay. The result is often the worst of both worlds: lower productivity and no serious control over the pace of change.

The better response is adaptation. That is a dull word, but the alternative is usually fantasy. Education systems need to pay more attention to the skills AI does not replace cleanly: judgment, communication, technical oversight, physical work, relationship management and the ability to work across messy real-world environments. People who can use AI tools without being replaceable by them will have an advantage.

Social safety nets will also matter more, not less. If AI makes the middle of the labour market thinner, then transition support becomes a practical necessity rather than a political slogan. Retraining sounds abstract until someone loses a job. Then it becomes the difference between recovery and drift. The point is not to preserve every old role. The point is to stop the transition from becoming a permanent scar.

Labour markets also need flexibility. The more barriers there are to moving between jobs, industries and regions, the harder it becomes to absorb technological shocks. That is why rigid systems tend to suffer most. They keep people attached to fading occupations for too long and make new ones harder to reach.

None of this is a grand theory. It is housekeeping for a changing economy. The big mistake is to imagine that policy can stop the weather. It cannot. It can, however, build better shelter.

The jobs AI is most likely to hit first

The public debate often gets lost because it treats “jobs” as one category. That is childish. Some jobs are far more exposed than others, and the differences matter.

The first to feel pressure are the roles built on repetitive information processing. Basic administrative work, routine drafting, first-pass research, standard customer support, simple image production and predictable content generation are all obvious targets. These are exactly the kinds of tasks where AI shines: high volume, low context and measurable output.

Jobs that depend on judgment in a messy setting are harder to replace. So are jobs where the worker’s presence is itself part of the value. That includes care work, skilled trades, face-to-face sales, field service, teaching, counselling and many forms of hospitality. AI may support those roles, but it will not remove the need for human contact as quickly as it removes paperwork.

The creator economy sits somewhere in the middle. It is exposed because digital content can be generated cheaply. But it is protected because audiences often buy personality, access and perceived authenticity. That is why platforms like OnlyFans are such a revealing case. The product is not merely content. It is intimacy sold at scale. Machines can imitate the outer form of that exchange. They struggle to replicate the belief that the exchange is real.

This does not mean creators are safe. Far from it. The most generic content will be crushed first. The creators who rely on volume alone may find themselves undercut by synthetic alternatives. But the ones who build a recognisable personal brand, maintain trust and offer something verifiably human may become more valuable. Scarcity can raise price.

That is the pattern to watch. AI does not simply replace workers. It sorts them. It favours the distinctive, the trusted and the physically present. It punishes the interchangeable. That is a harsh outcome, but it is more realistic than either utopia or apocalypse.

Why the noise around AI keeps missing the real story

The present debate is loud because it flatters both fear and vanity. Tech executives speak as if a new model can overturn civilisation by next quarter. Commentators speak as if every white-collar worker is already obsolete. Both positions make for easy headlines. Neither is much use.

Cowen’s intervention is useful precisely because it drags the argument back toward economics. It asks what AI actually costs, what the grid can carry, what consumers will pay for, where human presence still matters, and how quickly firms can change their operations. Those are the questions that decide employment. Not slogans.

The point is not that AI will leave the labour market unchanged. It plainly will not. It will reduce some tasks, expand others, and alter the bargaining power of workers in many fields. It will likely squeeze middle-tier work and reward people who can supervise, integrate or complement the technology. It may be especially disruptive in digital industries where the output itself can be generated at scale.

But the leap from disruption to mass unemployment is still too large. A technology has to pass through costs, constraints, institutions and preferences before it can remake the economy. That passage is slower than the press releases suggest. It is also slower than the panic.

The better way to understand AI is not as a clean replacement for human labour, but as a force that makes some forms of labour cheaper and more abundant while keeping other forms stubbornly human. That is less dramatic than the nightmare scenario, but it is closer to the truth.

And truth, in this debate, is not a luxury. It is the one thing that can still separate serious planning from expensive foolishness.