Why I believe the permanent underclass is coming

There is a countdown going around in technology circles. It reads: you have eighteen months to escape the permanent underclass. The number changes with whoever is posting it. The idea does not. Artificial intelligence will soon do most paid work, wages will stop being a way up, and whoever has not secured a foothold by then will stay where they are.

In April, Jasmine Sun's reported essay in The New York Times showed how far the idea has travelled inside the industry. It opens with a sentence that is hard to forget:

Most people I know in the A.I. industry think the median person is screwed, and they have no idea what to do about it.

I build software for a living, with the same tools that sit at the centre of this fear. I think the people in that sentence are right. Where I differ from the countdown is the number. Eighteen months is a sales pitch. My own estimate is fifteen years at the most. That is not a comfort. Fifteen years is shorter than a working life: a child starting school today will look for a first job inside it, and most people reading this will still be working when it closes. I trained as a statistician long enough to recognise a default when I see one, and this is one.

The argument

The careful version is older than the meme. Rudolf Laine's essay By default, capital will matter more than ever after AGI and The Intelligence Curse, written with Luke Drago, set it out. If machines can do human work, the value of a person's labour falls and the value of owning the machines rises. For most people in history, work has been the only ladder. Take the rungs away and the people at the bottom stay at the bottom, while the people who own things keep compounding.

I find this hard to argue with. The first reason is a chart.

The chart

Every year the Stanford AI Index compares the best models with a human baseline on a set of benchmarks. The dashed line at 100% is the human. What a statistician notices is not that the lines cross it. It is how fast they cross it, and how that speed keeps rising.

Image classification took about three years to climb from 90% of the human baseline to parity, from 2012 to 2015. Multitask language understanding took about five, from 2019 to 2024. PhD-level science questions went from under half of the expert baseline to above it in a single year. Autonomous software engineering went from about 65% to nearly 100% in one year. Computer use, an agent operating a desktop the way a person does, went from single digits to about three quarters, also in one year.

I know the objections, because I would raise them myself. The benchmarks are selected, and the ones that get solved are the ones that end up in charts like this. "Human baseline" means different humans on different tests, from crowd workers to domain experts. Test questions leak into training data. And a benchmark is not a job. All of this is true, and none of it explains away the newest lines. The early benchmarks measured seeing and reading. The newest measure work: writing software and operating a computer. Those are the lines with the steepest slopes, and they are the ones that matter for wages.

The chart also stops in 2025. Today, a model that leads the public rankings is often overtaken within two to four weeks.

Adoption is no longer slow

The strongest comfort used to be diffusion. Electricity took decades to change factories, because the factories had to be rebuilt around it. AI was supposed to be the same: capable in the lab, slow to reach real work.

Look at how fast the industry agreed on the plumbing. Anthropic published the Model Context Protocol, an open standard for connecting models to tools and data, in November 2024. Within months its rivals had adopted it, and thirteen months after launch it was handed to a new Linux Foundation body, with OpenAI, Google, Microsoft and AWS among its backers. Competitors do not agree on a standard that quickly unless they expect to need it. The link between the model and the systems where work actually happens is being built at the speed of software, not the speed of factories.

There is no next room to move into

The usual reply is history. In 1930 Keynes warned of "technological unemployment". In 1964 a group of scientists and economists sent President Johnson a memo, The Triple Revolution, predicting a permanent class with no work to do. Both were wrong, because people kept moving into new kinds of work.

But look at where they moved. Machines took the work of muscle, so people moved into routine thinking. Computers took routine thinking, so people moved into judgement, language, analysis and care. Every wave left a room the machines could not enter. This wave is aimed at that room itself. When the tool can read, write, reason and plan, there is no obvious next room. The economist Wassily Leontief made the comparison decades ago: the horse was not retrained when the tractor arrived. Its work simply ended.

"It never happened before" is a base rate. Base rates are useful until the mechanism changes, and this time the mechanism has changed.

Cheap to build is not the same as easy to climb

The hopeful view says the tools are rented by the month, at the same price for a student in Istanbul as for a fund in London, so everyone gets the same leverage. I see that leverage in my own work every day. Things that once needed a team, a budget and a year can now be made by one person.

That is good news for the one person. It is not good news for the team that was never hired. And when everyone can build, building stops being scarce. What stays scarce is what the builder does not get from a subscription: an audience, distribution, a brand people already trust, data, and the capital to wait while something finds its market. Those belong to whoever already has them. Building has become cheap. Being found has not.

The bottom rungs go first

Mobility is not about the top of the ladder. It is about the bottom. Entry-level roles are where people learned a trade while being paid for it: the junior developer, the trainee analyst, the assistant who became the manager. That is exactly the work the new tools do best, and it is the first work that stops being hired.

A ladder with its lower rungs missing still has a top. The people already on it are fine. Nobody new gets on. Do that for a generation and the result looks permanent, even though no single decision made it so.

Watch the median, not the mean

Statisticians learn early that an average can rise while most people fall behind. Gains from software tend to follow a power law: a few winners take most of the value. AI can raise total output enormously and still leave the median worker with a smaller share than before. Labour's share of income in many rich economies has been drifting down for decades. This technology pushes in the same direction, faster.

That is why the sentence in the Times names the median person. The median is the right statistic, and it is the one most likely to be left out of the headline numbers.

Cheap is a phase

Technology usually gets cheaper after it arrives, and AI seems to follow the rule: last year's capability now costs a fraction of what it did. But last year's capability is not what does the work. The frontier does, and the frontier has not become cheap. I read today's prices as a land grab. The largest labs are reported to be losing billions of dollars a year, paid for by investors who expect that money back.

Watch what the labs are doing with it. They are no longer only selling tools. They are buying companies: by the spring of 2026, OpenAI alone had made almost as many acquisitions as in the whole of 2025, in developer tools, testing, health and finance, and its rivals are doing the same. They are building products that do the work themselves, not products that help someone else do it.

When a lab sells the work, the small builder is no longer its customer. They are its competition, and there is little reason to keep the best intelligence cheap for a competitor. Today the student in Istanbul and the fund in London pay the same price. I do not expect that to last. When the land grab ends, I expect the frontier to become expensive, and access to it to gather around a few large companies and the people who own them. That is the real loss: not that the tools disappear, but that the best of them become something only the largest players can afford to run at scale.

Believing it is not the same as accepting it

I believe the permanent underclass is the default outcome. A default is what happens when nobody chooses otherwise. Sun's conclusion is the one I share: the production of a social underclass is a policy choice. That is exactly why it is worth saying out loud now, while the choice is still open, instead of after the ladder is gone.

For individuals, I do not think the answer is panic or a gamble before a deadline. It is a change in what you aim for:

• Own something. A product, a share, a body of work, an audience. Wages are what the machines compete with; ownership is what they work for.
• Build while building is cheap. The tools that threaten the job are, for now, the same tools that let one person make something of their own.
• Earn trust. When anyone can produce anything, people choose who they believe. Trust is slow to build and hard to copy.
• Take part in the argument. How the gains are shared will be decided by people. It is better decided early, and by many.

The countdown is wrong about one thing. There is no single moment when the door closes. It closes over years, not months, fifteen at the most by my estimate, one unhired junior at a time. That is the bad news. The good news is that something that closes slowly can still be held open.