The Jevons Paradox: Why AI Will Make Your Irreducible Human Skills 10x More Expensive
The Story You’re Being Sold
Everything you hear about AI and your job is framed in two ways: as inevitable, and as competition. The technology gets better, your job disappears, nothing anyone can do. You versus the machine, and the machine is faster, cheaper, and never sleeps.
Both frames are wrong. Or at best, they represent one way of seeing things, and they happen to be the way that benefits the people selling the technology.
When Mustafa Suleiman, CEO of Microsoft AI, recently argued that AI is on track to reach human-level performance on most professional tasks done at a computer, and that many of those tasks could be fully automated within twelve to eighteen months, he wasn’t issuing a warning. He was making a pitch. His company and other major AI firms have committed, and are on track, to invest on the order of hundreds of billions, approaching three-quarters of a trillion dollars, in AI infrastructure over a few short years. That money needs a return. One way that math works for investors is if AI replaces human labor at massive scale, or generates equally massive new revenue streams.
So when industry leaders tell you your job might be gone in eighteen months, understand the context. They’re telling their investors the investment will pay off.
The Loop Nobody’s Talking About
Here’s the question nobody seems willing to walk through: the economy is a loop. Companies pay workers. Workers buy things. Revenue comes back. Automate away the workers and the customers disappear. Who exactly are you selling AI to then? The other AIs?
Henry Ford figured this out a century ago. He paid his workers enough to buy the cars they built. It wasn’t charity, it was a minimum viable economy.
If every company replaces its people with the same AI models, what is any company actually worth? Your talent was your moat, your institutional knowledge, the judgment your people built over decades. The argument that you don’t need the people, that you just need to capture what they knew and train it into an AI, misses something fundamental. What you’ve captured is a snapshot, a photograph of institutional knowledge taken on the day you extracted it. The world moves. The snapshot does not. The people who would have adapted to the next disruption you didn’t see coming are gone.
Institutional knowledge isn’t a stockpile you can extract and keep. It’s a river. Cut off the source and it dries up.
The 75 Million Who Read the “Something Big Is Happening” Article!
Matt Schumer, an AI startup CEO, recently published a piece comparing the present moment to February 2020, right before COVID hit. He said AI was replacing him at his own job and was coming for everyone else’s next. His advice? Learn to use AI and tidy up your finances. That’s it, a general diagnosis with no options.
By his account, that piece went extremely viral, reportedly reaching on the order of 75 million readers. And research on social movements suggests it can take as little as a few percent of a population, often cited around 3.5%, actively mobilizing to drive large-scale change. If even a fraction of those readers felt the mix of anger, fear, and skepticism that many of us did, this moment could actually matter. Not necessarily because he’s right, but because people are finally paying attention.
His argument rides on one thing: code. AI got exceptionally good at writing software. But software is one of the most AI-friendly domains, clear right answers, automated testing, a machine that can check its own work. Generalizing from that to all knowledge work is like watching a calculator beat you at math and deciding it’s about to outperform you at parenting.
Most real work is messy. A contract requires strategic context. A medical diagnosis runs on incomplete information. A policy decision weighs values that can’t be optimized. Early deployments already show a pattern: AI can accelerate routine tasks, but the highest-stakes parts of work increasingly revolve around human judgment, responsibility, and context.
There is growing evidence that, in many settings, AI is increasing the need for human judgment rather than erasing it, because it expands what can be attempted and raises the stakes of final decisions instead of making humans irrelevant.
What You’d Have to Believe
The argument that coding breakthroughs generalize to all knowledge work, that software mastery becomes universal mastery because software touches everything, is genuinely interesting. But “software touches everything” doesn’t mean everything reduces to software. Water touches everything in biology. That doesn’t mean mastering hydraulics makes you a neuroscientist.
To believe the full generalization story, you’d have to believe several things simultaneously.
You’d need to believe in real-time continual learning, when current AI systems do not update their core parameters from experience the way humans do, they are effectively frozen between training runs. Real knowledge work requires a professional who learned something last Tuesday in that specific room with that specific client, and carries it forward into the next decision. That’s not just retrieval. It’s a fundamentally different cognitive act.
You’d have to believe AI can reach what complexity theorist Stuart Kauffman calls the “adjacent possible”, but his entire point is that it’s locally determined, calibrated to human bodies and minds built over decades of lived experience.
You’d need something close to 100% reliability. In a complex system, errors compound. Even if each step in a fifteen-step process were 95% reliable, the chance that the entire chain is error-free is only about 46%. Add more steps or slightly lower per-step reliability, and the probability of a clean outcome plunges. For anything with real stakes, the math gets brutal fast.
You’d have to believe that tacit knowledge, what we know but can’t articulate, can be fully captured. If tacit knowledge could be fully formalized, we’d already see much more of it sitting cleanly in datasets; the fact that we don’t is a sign that formalization is partial at best. The claim that software mastery reaches everything is, in effect, the claim that what couldn’t be written down doesn’t matter. There’s a great deal of evidence from medicine, management, and skilled trades that it does.
And you’d have to believe in either full physical embodiment or, more uncomfortably, extremely deep surveillance, a therapist reading micro-expressions, a surgeon feeling tissue resistance, a teacher noticing a child who’s gone quiet. Either AI gets a body that can do all of that, or it needs to watch us so closely that it no longer needs one. Neither is here today at any general scale. And the second comes with a different set of problems entirely.
Accountability: The Line AI Cannot Cross
Here’s the part nobody discusses when they say AI will replace your job. AI can make a recommendation. What it cannot do is be accountable for one.
When a doctor recommends treatment, they put their license on the line. When a lawyer gives advice, they’re personally liable and bound by professional standards. When a manager makes a call, they own the outcome in their organization. AI can generate the recommendation, but it can’t stand in court, lose a license, or be fired.
As humans, we expect someone to own it. We need someone to look us in the eye and say: why did you decide this? That’s why human judgment isn’t going away, not because AI isn’t smart enough, but because accountability requires a person.
Across thousands of professionals studied over several years, a consistent pattern emerges: the people thriving with AI are those who know exactly what they’re accountable for and how they show up for others. They use AI to extend their reach, not to abdicate responsibility.
The Dust Bowl Warning
In the 1920s, American farmers pushed into the Great Plains. They plowed under native prairie grasses, diverse, deep-rooted, evolved over thousands of years, and replaced them with single cash crops. For a few years, yields were incredible. Then the drought hit. Without those complex root systems holding the soil together, the earth simply blew away.
What’s happening in some organizations right now follows a similar pattern. AI leaders are plowing under the root systems of human work, mentorship, the judgment that builds through friction, institutional knowledge that lives in people rather than documents. Short-term, it looks like efficiency: fewer headcount, cleaner margins, better quarterly numbers. But our economy, like that prairie soil, depends on human complexity that took decades to grow.
The Dust Bowl took years of stress to reveal how fragile monoculture was. An AI monoculture could fracture even faster, because so few organizations are actively monitoring what’s being lost as they rush to automate.
Baumol’s Cost Disease and the New Price of Human Work
Your hairstylist takes the same amount of time to cut your hair as fifty years ago. No technology has made that faster. Yet it costs many times what it did in 1975. Why? Because everything around your stylist got more productive. This is Baumol’s cost disease: when technology makes some sectors wildly productive, sectors that can’t automate much get more expensive. Not because they got faster, but because wages rose elsewhere.
For two hundred years, this was easy to see. The work that stubbornly needed a human was usually physical. You couldn’t download a haircut. But AI changes the equation because it’s the first automation technology pushing deeply into territory that used to be protected by Baumol, education, healthcare, legal advice, therapy, creative work.
So does the principle still hold? It does, but the line has moved. AI can teach calculus, and in many cases will tutor more consistently than a typical human tutor. But the teacher who notices your child stopped making eye contact three weeks ago and calls you about it, that’s a different kind of work entirely. And that work is repricing upward.
The question that used to be simple, “can a machine do this job?”, is now harder and more important: which specific part of this work stubbornly needs a human, even when AI can do everything around it? Next time you finish a piece of work, look at what you actually did. How much was process and how much was judgment? The process part is getting cheaper by the month. The judgment part is your price going up.
Weak Links: Why the Bottleneck Becomes the Whole Game
AI leaders talk about a “country of geniuses in a data center”, billions of AI models doing every cognitive task humans can do, but cheaper and faster. There’s an economic principle almost nobody in that conversation is using: a system is only as productive as its least productive essential part.
Think about a flight. You can automate booking, check-in, baggage handling, navigation, even most of the flying. Then the plane lands and needs a gate, a ground crew, and three hundred people shuffling down a narrow aisle. Someone needs to de-ice the wings at 5 a.m. in February. Automate everything in the air, and the bottleneck simply moves to the ground. You don’t eliminate the constraint, you relocate it.
In one influential model, a Stanford economist shows that if software represents about 2% of the economy, then even making software tasks infinitely productive only raises GDP by around 2%. In more aggressive scenarios, automating essentially all cognitive tasks yields on the order of a 50% increase in output, a big number, but nowhere near infinity, and it comes at the price of replacing nearly all human thinking. The transformation doesn’t look like everything getting replaced at once. It looks like value concentrating in the tasks that resist automation.
The weak links become the expensive links. The bottleneck becomes the whole game.
Hard Parts vs. Easy Parts: The Question That Actually Matters
Taxi drivers and accountants both got automated. One group got poorer. The other got richer. The reason why is perhaps the most important idea for understanding what AI is about to do to your career.
Before Uber, London cabbies spent years memorizing 25,000 streets, “the Knowledge.” That expertise was the entire job. Then GPS automated the expert part, the hard part. Suddenly anyone with a car and a phone could do the job. Employment in ride services surged. Wages stayed flat or fell relative to the old, restricted system.
Now look at accountants. Computers automated the routine part, data entry, bookkeeping, repetitive calculations. The easy part. What remained was complex analytical work requiring more expertise, not less. Wages went up. The job became more specialized and more valuable.
Same dynamic, technology automates part of a job, with completely opposite outcomes. The difference: whether the technology took the hard parts or the easy parts. The right question isn’t “will AI take my job?” It’s: in my job, is AI taking the hard parts or the easy parts?
If AI is consuming the drudge work and leaving you with the judgment calls, the economics are in your favor. If it’s eating the expertise and leaving you with the administration, the situation is different entirely.
Saturation, Jevons, and the Disappearing Middle
Some creative workers are losing their jobs right now, copywriters, translators, illustrators. Some of this is short-sighted on the part of companies doing it, but some of it isn’t.
When the cost of producing something drops to nearly zero, demand doesn’t go to infinity. It saturates. There are only so many blog posts you can read, only so many product descriptions anyone needs. Think about those generic ads in your feed, “Oregon drivers delighted by this new insurance change”, with their AI-generated stock images. Nobody ever really cared about those images. They were filler, filling a rectangle. Filling a rectangle is not a job that survives when a machine can fill it for free.
AI didn’t devalue that work. It revealed what the market actually valued it at.
But saturation and explosion are happening simultaneously. While commodity content saturates, AI-assisted “vibe coding” is blowing open a universe of problems that software never reached before, problems too local, too specific, too contextual for any commercial product to justify building. Economists call this a version of Jevons paradox: when a resource gets dramatically cheaper, consumption often expands into territory that was always there but never reachable.
If your work is commodity, replicable, indistinguishable from what a machine produces, that market is collapsing and isn’t coming back. But if your work represents twenty years of frontline knowledge about a specific hard problem that nobody ever built software for, that market is just opening.
The Loop That Could Break
The saturation story isn’t just about careers, it’s an economic one. The economy is a loop of people buying from each other. The plumber fixes the tutor’s sink. The tutor teaches the plumber’s kid. The accountant does both their taxes. Each person earns because they’re someone else’s customer.
Under pure substitution, this loop doesn’t just slow, it structurally changes. And here’s what should concern everyone, including the people building AI: you can build the most productive company in history, but if large parts of the population fall out of the income loop, you shrink your own market.
The forces of complementarity, task reshaping, the persistence of judgment, the way automation can increase the value of remaining human work, push against a simple collapse narrative. The loop doesn’t break automatically, but it depends on how organizations interpret automation. If jobs are treated as task lists that disappear one by one, substitution dominates. If jobs are understood as bundles where automation removes commodity elements and amplifies human ones, value redistributes rather than vanishes.
Efficiency Has a Ceiling. Capability Expansion Doesn’t.
Workers say they feel more productive. CEOs say gains are uneven. Economists say it isn’t yet showing up cleanly in the aggregate data. These accounts can all coexist because one word, “productivity”, is describing several different dynamics.
Many conversations about AI are framed through the efficiency lens: how quickly existing work can be completed. But there’s another process unfolding, the expansion of the possibility space. New capabilities allow us to attempt work that previously felt out of reach. Drug discovery is a good example: AI systems can now explore chemical space at previously inaccessible scale, but that capability exists long before new medicines move through trials, regulation, manufacturing, and markets.
Look for moments this week when AI changed what you were willing to attempt. Find an idea you could move from abstract to testable. Find a problem that became approachable because you used AI. Efficiency has a ceiling. Capability expansion has none. And there are good reasons to believe that skilled, curious humans will choose the trajectories that AI can unlock.
The Country of Geniuses, In Us, Not a Data Center
Dario Amodei’s image of AI as “a country of geniuses in a data center” is revealing for what it implies. It puts the genius in the machine, concentrates intelligence in one place, and sets up the entire conversation as: how do we deal with the fact that the genius is over there and not in us?
Imagine instead: what if we could make every worker a genius at their job? Same technology, completely different world. In that framing, you’d invest in people, build capability upward rather than replace it. And the economics back this up, the most important knowledge in an economy is rarely centralized. It’s local. The account manager who senses which client is about to leave before any metric picks it up. The product manager who knows which feature requests are really about a completely different problem. That knowledge doesn’t sit neatly in any public dataset.
A country of geniuses in a data center is a monoculture, one or a few models, similar outputs, similar answers to everyone. A country of actual people with different knowledge and different contexts, each augmented by AI, that’s where competitive advantage comes from.
Concentrate intelligence and you flatten it. Distribute it and you multiply it. We’ve seen versions of this in economics, in complexity science, and in history: monocultures look efficient right up until they collapse.
So What Do You Actually Do?
Yes, you should be using AI. That part is real. But how you use it matters more than whether you use it. And that’s a question about agency, authorship, and purpose, not about panic.
Start by separating process from judgment in your own work. Offload as much commodity process as you can to AI, drafts, boilerplate, basic analysis, and reinvest that time in the parts of your work that stubbornly require a human: interpretation, relationships, responsibility, creative leaps. Move yourself closer to the bottlenecks: diagnosing problems, coordinating across functions, being the person who owns the decision.
A long history of economic thinking suggests that when new technologies complement human skills rather than substitute for them, the combined system is more valuable than either alone. The story you’re being sold is that AI and human intelligence are in zero-sum competition. The quieter story is that the pairing can be super-additive, but only if we choose to build it that way.
The real story isn’t inevitability, it’s complexity. There are decisions, incentives, institutions, and economics that shape what actually happens, and the outcome depends on choices we haven’t made yet. The real dynamic isn’t simple competition, it’s cooperation. The question isn’t whether AI will replace you. The question is: what becomes possible when human judgment and artificial intelligence work together?
That question is worth a lot more than fear.
In my Notion, I have a section called “My Articles”, and my articles are there, but I don’t know how I captured it in a way that, when you open it, there is a header and the picture. Everything. How did I import them into Notion? I forgot what was the method.