The Boiling Frog Economy.

The first mistake is expecting the AI crisis to look dramatic.

People imagine the labour market breaking like a bridge: one sudden collapse, one shocking unemployment number, one obvious moment where machines replace humans and everyone finally agrees that the world has changed.

But that is probably not how it happens.

The more dangerous version is quieter. The economy keeps moving. The stock market rises. Companies talk about productivity. Workers use AI tools and feel faster. Founders build smaller teams. Senior employees produce more. Managers delay hiring juniors. Freelancers lower their prices. Graduates send more applications and hear less back. Nothing looks like a collapse. Not yet.

The water gets warmer.

That is the Boiling Frog Economy: a world where AI does not immediately destroy the job market, but slowly weakens the ladder that ordinary people climb.

For decades, the path into the middle class was built on imperfect but functional steps. A person went to university, found an entry-level role, did repetitive work, learned from mistakes, gained judgment, became useful, moved up, saved money, and eventually built some security. The work at the bottom was often boring: writing drafts, fixing bugs, reviewing documents, making spreadsheets, answering customers, preparing slides, summarising research. But those tasks had a hidden purpose. They trained people.

AI is now strongest at exactly that layer.

It writes the first draft. It builds the basic app. It summarises the document. It answers the support ticket. It produces the test case. It edits the marketing copy. It scans the contract. It creates the spreadsheet model. It does the thing a junior used to do badly, slowly, and educationally.

That is the uncomfortable truth. AI may not replace the expert first. It may replace the training ground that creates the next expert.

---

I. A job is not one thing

Most public discussion asks, “Will AI take our jobs?” That question is too crude.

A job is not one thing. A software engineer does not only write code. A lawyer does not only draft contracts. A consultant does not only make slides. A marketer does not only write copy. A researcher does not only summarise papers. Jobs are bundles of tasks, judgment, trust, communication, context, and accountability.

AI does not attack the whole bundle evenly. It attacks the parts that are easiest to describe, digitise, repeat, and verify.

This is why the economy can look stable while the structure underneath changes. A company may not fire an entire department. It may simply avoid hiring five juniors next year. It may replace contractors with AI tools. It may expect each employee to produce twice as much. It may keep headcount flat while revenue grows. It may hire one senior operator instead of a team. The official unemployment rate may not scream. But the ladder quietly narrows.

That is why AI exposure is not the same as job loss. Exposure means AI can perform some of the tasks inside a job. Job loss happens later, after companies adopt the technology, redesign workflows, trust the output, change hiring plans, and decide that fewer humans are needed. The IMF estimates that around 40% of global employment is exposed to AI, rising to about 60% in advanced economies. In those advanced economies, roughly half of exposed jobs may benefit from AI, while the other half may face lower labour demand, lower wages, reduced hiring, or disappearance in extreme cases.¹

That distinction matters. The future may not be a clean story of humans versus machines. It may be a messier story: some humans become far more powerful with AI, while others become easier to replace.

---

II. The first casualty is not work. It is apprenticeship.

The old economy had a rough bargain. Young workers were not very good yet, but companies hired them because they were cheap, trainable, and necessary. Juniors did the low-level work. Seniors reviewed it. Over time, the juniors absorbed taste, judgment, habits, and context. Eventually, they became the people who could handle ambiguity.

AI damages that bargain.

If a senior accountant can use AI to prepare first-pass reports, why hire as many junior analysts? If a senior engineer can generate boilerplate, tests and documentation, why hire as many junior developers? If a lawyer can use AI to review standard contracts, why bring in as many junior associates? If a marketing lead can generate 20 campaign drafts in an afternoon, why pay a beginner to do it?

The junior was never hired because the first draft was perfect. The junior was hired because the first draft had to come from somewhere. Now it comes from the machine.

This is why the strongest early warning signs are appearing around young workers. A Stanford Digital Economy Lab study found that early-career workers aged 22 to 25 in highly AI-exposed occupations experienced a 16% relative employment decline, while experienced workers in the same occupations remained more stable or continued to grow.² It is one of the most important signals in the entire AI labour debate. It suggests AI may not flatten the labour market all at once. It may hollow it from the bottom.

The senior survives because they know what good looks like. The junior struggles because the tasks that would have taught them what good looks like are now done by the machine. The same tool that raises the ceiling also lowers the staircase.

---

III. The strongest argument against this thesis

This view deserves its sharpest opponent. The optimistic case is neither silly nor naive, and the people who hold it can point at history.

The argument runs roughly like this. Every previous wave of technology eventually created more work than it destroyed. Tractors did not produce permanent agricultural unemployment; they freed labour for industries that did not yet exist. Personal computers did not erase the office; they enlarged it. The internet vaporised plenty of intermediaries and minted entire industries no one had drawn on a roadmap. By that pattern, AI should be another lever in the long story of capital making humans more productive, not a wall that humans cannot climb. If it lowers the cost of starting a company, a careful operator with good taste and a model subscription can compete with a whole agency. If it raises the floor for novices, a first-year analyst with AI may already match a third-year analyst without one. If frontier tools become commoditised, the small operator gains as much as the platform giant.

Some of that will be true. Probably more of it than the pessimists allow. New categories of work will appear. The cost curve of building things will keep falling. Individuals with judgment and access will do work that used to require firms.

The hard question is not whether new jobs appear. It is whether they appear in time, in the right places, and for the right people. New industries usually take a decade or longer to absorb the workers a previous wave displaced. They tend to need different skills, often skills that come from training the displaced cannot easily afford. They concentrate geographically, with a handful of cities collecting most of the upside while older labour markets contract. The historical comfort that “technology eventually creates more jobs than it destroys” rests on a quieter assumption: that the new jobs were within reach of ordinary people. That is the assumption AI most directly tests.

So the real question is not “Will there be jobs?” It is whether the path from no experience to high-leverage work remains realistic for someone without capital, network or geography on their side. If that path stays open, the optimists are right and most of the rest of this essay is wrong. If it narrows, technology can keep producing jobs in aggregate while still producing an economy where ordinary people lose ground.

The wager of this essay is that the second version is the one we are entering.

---

IV. The broken ladder

A healthy economy does not only need jobs. It needs pathways.

It needs ways for someone with little money, little network and little experience to move into better work. The ladder matters because most people do not start with capital. They start with labour. Their labour is supposed to become skill. Skill is supposed to become income. Income is supposed to become savings. Savings are supposed to become assets. Assets are supposed to become security.

If AI weakens that progression, the consequences are larger than unemployment.

Brookings has warned that AI could reshape career pathways from lower-wage Gateway jobs into higher-paying Destination jobs. Its analysis found that only 51% of pathways between Gateway and Destination occupations are not highly exposed to AI.³ In plain English: nearly half of the routes people use to climb into better jobs may be exposed to disruption.

That is the deeper social risk. A person may still find work. But the work may not lead anywhere. They may be employed but stuck. Productive but poor. Busy but unable to accumulate. Educated but unable to enter the first serious role. Surrounded by software that makes everything faster except their own escape from insecurity.

This is how a society becomes more rigid without looking broken. There is still growth and there are still jobs. Fewer people can cross from the bottom into the high-leverage layer. The ladder does not collapse with noise. It becomes thinner, more selective, and harder to find.

---

V. Labour versus ownership

The deepest AI question is not technical. It is economic.

If AI produces enormous value, who captures it?

McKinsey estimated that generative AI could generate between US$2.6 trillion and US$4.4 trillion in annual value across industries.⁴ That is not a small software upgrade. That is a global redistribution event waiting to happen. But value creation is not the same as value distribution.

The people most likely to capture the gains are those who own scarce assets: models, chips, cloud infrastructure, data centres, proprietary datasets, distribution channels, software platforms, customer relationships, regulated workflows, and companies. The people most exposed are those who sell undifferentiated labour into tasks that AI can reproduce cheaply.

This is the capital problem.

In the old economy, labour still had bargaining power because companies needed people to produce output. In the AI economy, output increasingly comes from a combination of human direction, software, compute, and capital. If a firm can produce more with fewer people, the value of labour can fall even while total productivity rises.

That means GDP can go up while workers feel poorer. Corporate margins can improve while entry-level hiring declines. Founders can build leaner companies while graduates struggle to get interviews. A country can become more productive while ordinary people lose leverage.

That is the part people often miss. AI does not need to make humans useless to change the balance of power. It only needs to make many forms of labour less scarce. When labour is less scarce, ownership matters more. If you want to see where this ownership stack actually lives (energy, chips, infrastructure, models, applications), that is the five-layer geometry I mapped in the AI Map.

---

VI. Software engineering is the preview

Software is where this future is easiest to see.

AI can already build landing pages, generate CRUD apps, write API endpoints, create basic tests, explain code, refactor components, and implement standard features. This does not mean software engineers disappear. It means the value of “I can write code” gets compressed.

The valuable engineer becomes the person who knows what should be built, how the system should work, what tradeoffs matter, what can fail, what is secure, what is maintainable, and whether the AI-generated output is actually correct. That is a very different job.

The weak version of software work is task execution: build this page, write this function, fix this bug. The strong version is system ownership: understand the customer, design the architecture, protect the data, choose the right tradeoff, verify the output, and carry responsibility when it breaks. AI attacks the first version much harder than the second.

This is why junior developers are under pressure. Not because AI can replace every serious engineer, but because it can perform many of the bounded tasks that used to justify hiring beginners.

The same pattern spreads beyond software. In law, AI attacks document review and first-pass drafting. In consulting, research summaries and slide creation. In marketing, basic copy and campaign variations. In finance, spreadsheet work and report generation. In customer support, scripted responses. In administration, scheduling, data entry and routine communication. The pattern is identical everywhere: AI compresses the bottom layer first.

---

VII. Why the crisis may not look like a crisis

A major mistake is assuming that if AI were truly disruptive, unemployment would already be exploding. That is not how structural change always appears.

Companies do not need to announce, “We replaced juniors with AI.” They can simply hire fewer of them. They can reduce internship programmes. They can avoid replacing people who leave. They can cut agencies. They can give existing staff AI tools and raise expectations. They can freeze headcount while revenue grows.

This creates a quiet labour-market shift. The pain is concentrated among people trying to enter, not necessarily people already inside.

That is why early evidence can look contradictory. Anthropic’s labour-market research finds limited evidence so far that AI has broadly affected employment.⁵ The Bank of Canada has also said there is no current evidence of large-scale AI-driven job loss, while still expecting AI to transform many tasks.⁶

Both things can be true. AI may not have caused mass unemployment yet. AI may still be weakening the entry layer.

---

VIII. The timeline of the boiling

If the optimistic case were obviously right, the unemployment data would already look like a problem. It does not. Goldman Sachs projects roughly 6 to 7 percent of workers displaced over a ten-year transition, with a wider band of 3 to 14 percent depending on adoption speed.⁷ The World Economic Forum forecasts 170 million new roles and 92 million displaced by 2030, a net positive of 78 million jobs.⁸ Aggregate numbers like that are the favourite tools of the optimist, and they are not wrong. They simply do not describe what individual people experience while the transition runs.

The transition has its own shape. Here is the version I think is most plausible.

There are alternative versions of this story. One is a financial-shock scenario, in which AI valuations and infrastructure spending outrun real revenue, the bubble corrects, and firms use the downturn to cut headcount they had hesitated to cut before. Another, raised most often outside rich countries, is the collapse of outsourced digital labour as a development pathway. Both are possible. Neither replaces the central forecast, which is uneven compression rather than sudden disaster.

People do not move between occupations the way tokens move between buckets. They have rent, debt, families, visas, geography and limited time. Economists can write that “new jobs will be created” and be entirely correct, while the displaced admin worker does not become an AI infrastructure engineer, the squeezed junior developer does not become a senior architect, and the call-centre worker does not move into cybersecurity. Aggregate gain coexists with individual immobility. That is what the timeline is really about.

---

IX. What the Boiling Frog Economy feels like

It does not feel like the end of the world.

It feels like applying for 300 jobs and hearing nothing. It feels like watching companies announce record productivity with smaller teams. It feels like freelance rates collapsing because clients believe AI can do the first 80%. It feels like every job requiring experience, while fewer places are willing to give it. It feels like being surrounded by progress that somehow does not make your life easier.

This is why the frog metaphor works.

The danger is gradual adaptation to worsening conditions. People adjust their expectations downward. They accept weaker wages. They accept unpaid trials. They accept constant reskilling. They accept that entry-level jobs require senior-level portfolios. They accept that buying a house is impossible. They accept that ownership is for someone else.

And because the decline is gradual, it becomes normal.

That is the real threat. Not that the system stops. That people stop believing they can climb inside it.

---

X. What survives

The conclusion is not that humans are doomed. The conclusion is that the safe strategy changes.

In the AI economy, the worst position is to be a generic seller of replaceable tasks. The best position is to own or control something that AI makes more valuable. That can mean owning a company, a product, a customer relationship, a trusted brand, a niche workflow, a dataset, a distribution channel, a regulated process, a community, or a high-trust service layer.

The future belongs less to people who can merely do tasks and more to people who can define problems, judge outputs, build trust, coordinate systems, understand customers, and capture value. The most valuable human will not be the one who refuses AI. It will be the one who can direct AI toward a real outcome and take responsibility for the result.

That is why judgment matters more, not less.

When output becomes cheap, taste becomes valuable. When code becomes cheap, architecture becomes valuable. When labour becomes less scarce, ownership becomes decisive.

The Aristotle move I wrote about in an earlier essay, learning how to think before being given anything to think about, matters more in this world, not less. It is the only capability AI cannot hand you for free.

---

The final thesis

AI will not necessarily break the job market in one dramatic moment. It may do something more subtle. Routine work gets cheaper. Junior roles thin out. Seniors become more productive. Teams contract. Wages in exposed work weaken. Income concentrates toward owners. The route up narrows. Years later, people look up and realise the ladder has been burning the whole time.

That is the Boiling Frog Economy. Not the sudden end of work, but the slow loss of worker leverage. Not machines replacing everyone, but a system where the people who own the models, platforms, data, compute and distribution gain power faster than everyone else can adapt.

---

Ownership is the new literacy. The frog cannot vote on the temperature of the water. It can only step out of the pot.

This is Essay No. 003. The topics: intelligence, AI, systems, knowledge, and the questions underneath the questions everyone else is asking. If you read this far and disagreed with any part of it, write to me. I read everything.