For roughly 10,000 years, every economy we built needed human beings to do things. To plant. To build. To calculate, analyze, write, and decide.
Our ability to do those things gave us leverage. It helped determine why we were hired, what we were paid, and how much bargaining power we had.
So what happens to that leverage when a machine can think what you can think?
Hold that question. My thesis I want to offer is that the marginal cost of intelligence is moving toward zero. The consequences are more complicated than the boosters suggest and more urgent than the skeptics allow. They matter most for education, because schools are preparing students to enter an economy in which cognitive ability may no longer be scarce.
Kill the Comfortable Idea
I want to start by killing a comfortable idea: that AI is simply another tool.
In that view, AI belongs on a shelf beside the printing press, the steam engine, electricity, the computer, and the Internet. Each of those technologies changed what humans could do. AI is the first one built to change what humans need to do at all.
That is a difference of kind, not merely a difference of degree. Four properties, taken together, explain why.
First, AI compresses the cost of cognition, not only the cost of labor. Analysis, writing, synthesis, translation, and research are among the things we have historically paid educated people to do. AI can now produce some of that work at extraordinary speed and falling cost.
Second, AI scales with compute rather than headcount. Historically, adding capacity meant adding people, machinery, or both, and each carried a cost. AI systems can expand output without a corresponding increase in the number of people doing the work.
Third, AI arrives across domains at once. The steam engine moved through manufacturing, transportation, and agriculture over decades. AI is being deployed in medicine, law, software, scientific research, and education simultaneously.
Fourth, AI does not only analyze. It generates, creates, and interprets. The quality remains uneven, and the boundaries matter, but it now performs work we have long treated as distinctly human.
No single property is unprecedented. The combination is.
“Every earlier technology changed what we could do. This one changes what we need to do at all.”
Take the Skeptical Case Seriously
There is an honest counterargument, and it deserves more than a footnote.
MIT economist Daron Acemoglu argues that AI’s macroeconomic effects may be meaningful but substantially smaller and slower than prominent forecasts suggest. His analysis estimates no more than a 0.66% increase in total factor productivity over 10 years. It also concludes that roughly 5% of tasks may be profitably automated in the near term.
He may be right. Hype is real, and technical capability does not automatically become reliable, economical deployment.
But notice what the disagreement is about. It is about timeline and magnitude, not about whether AI is a new kind of capability. In the short run, hype can outrun what organizations can deploy. Over the long run, the capability can still reshape the structure of work. We should update as evidence comes in, without mistaking uncertainty about speed for evidence that the shift is ordinary.
Fourteen Years from Ignition to Action
The story of AI adoption is not the list of dates. It is how quickly the gaps between them are closing.
In 2012, AlexNet, a neural network developed by Krizhevsky, Sutskever, and Hinton at the University of Toronto, won an image-recognition competition by more than ten percentage points over its closest competitor. Deep learning moved rapidly from the margins to the center of AI. That was the ignition. Five years later, researchers at Google introduced the transformer architecture behind today’s large language models. Three years after that, OpenAI’s GPT-3, with roughly 175 billion parameters, showed it could perform tasks for which it had not been explicitly trained. Today, systems are beginning to act rather than only answer, reliably executing defined multistep workflows inside controlled environments. Fourteen years separate ignition from this early era of autonomous agents.
I have seen the shift in my own work. I built a financial analytics system by describing what I wanted in plain English. I did not write a line of its code. I would not claim that AI makes everyone instantly capable of building such a system. The narrower point is more consequential: the bottleneck is moving from building the thing to knowing what to build and judging whether the result is any good.
Then the technology spread. ChatGPT reached 100 million users in about two months, compared with roughly nine months for TikTok. Among U.S. adults ages 18 to 64, generative AI adoption rose from 45% in late 2024 to 62% by May 2026. Measured at the same point after launch, that is faster than either the personal computer or the Internet.
A district strategic plan may cover three to five years. On this timeline, the technology can change substantially, possibly more than once, inside a single planning cycle. A policy written once and filed away is not a strategy.
“The adoption, in some ways, has already happened. The only question is whether it happens with your guidance or without it.”
The Missing First Rung
For educators preparing the next generation of workers, this shift becomes concrete at the beginning of a career.
Bloomberg estimates cited by the World Economic Forum suggest that AI could automate 67% of the tasks performed by sales representatives, compared with 21% of the tasks performed by the sales managers directly above them.6 The gap is not an accident of seniority. It reflects what the jobs are made of.
Junior roles often contain more research, drafting, summarizing, and first-pass analysis, which are among the tasks AI performs most readily. The role above them contains more judgment, relationships, context, and accountability for decisions. Exposure is distributed by task, not by title.
This is not only a projection. An August 2026 update from the Stanford Digital Economy Lab found that employment for workers ages 22 to 25 in highly AI-exposed occupations was about 19% below where it would have been had it kept pace with employment among less-exposed peers. 7
The junior role is more than a job. It is where judgment gets built.
A young employee does routine work in exchange for proximity to people who know things. They watch how an experienced colleague frames a problem, spots an exception, handles ambiguity, and takes responsibility for a decision. A few years later, that employee knows things too.
If the routine work disappears, that exchange breaks. The next rung on the ladder is still there, but it may sit too high to reach from the ground.
Schools stand at the other end of this transition. Much of our education model assumes that a first job will be waiting to receive the graduate and continue the process of building judgment. A 22-year-old leaving school needs more than a credential or a portfolio. They need somewhere to learn the craft.
If the first rung changes or disappears, preparing students to reach it more efficiently is not enough. We have to ask what replaces the learning that used to happen there.
Same Technology, Opposite Outcomes
The evidence does not point to a single predetermined result.
The same Stanford research found employment declines concentrated in occupations where AI substitutes for human tasks. Employment was flat or rising in occupations where AI complements what workers do. The researchers separated the two by examining whether large-language-model queries appeared to replace occupational tasks or help workers carry them out.
Same technology. Opposite outcomes.
The difference is not simply what the technology can do. It is how an employer chooses to deploy it: to make people more capable, or to replace the work they do. Employers make that choice. Policy and education help shape the conditions around it.
Schools do not control the employer’s decision. They do influence what graduates bring into the workplace, and therefore whether augmentation is a realistic option. A graduate who can frame problems, challenge outputs, navigate uncertainty, and exercise judgment offers an employer something different from a worker prepared mainly to execute repeatable cognitive tasks.
That is a narrower claim than saying education determines the outcome. It is also one the evidence can support.
What Becomes Scarce?
If intelligence becomes abundant, value moves toward what remains scarce. Three things stand out.
Discernment, because an AI system can be confidently wrong, and someone has to know when. That includes the ethical version of the question: not only whether an output is correct, but whether what it enables should be done at all.
Purpose, because executing quickly is worth little if the goal is poorly chosen. So is originality in the real sense: not producing more content, but originating an idea without a clear precedent.
Systems thinking, because decisions produce second- and third-order effects that do not fit inside a single prompt, and new situations rarely arrive labeled with the framework that explains them.
None of these is a standalone subject to add to the schedule. They have to be practiced inside history, biology, literature, mathematics, and the rest of the curriculum.
So, back to the question I asked you to hold. What happens to our leverage when a machine can think what we can think?
It doesn’t disappear. It moves: toward judgment, and toward whoever decides how these systems get used. For 10,000 years, the economy needed human cognition, and that need gave people a seat at the table. As the cost of cognition falls, that arrangement becomes negotiable.
Technology does not settle the negotiation. The choices institutions make now, and the skills schools help students build, will determine what they bring when they take that seat.
- Daron Acemoglu, The Simple Macroeconomics of AI, May 2024.
- Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton, ImageNet Classification with Deep Convolutional Neural Networks.
- Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin, Attention Is All You Need, 2017.
- Language Models Are Few-Shot Learners, 2020.
- Alexander Bick, Adam Blandin, and David Deming, “The Rapid Adoption of Generative AI,” St. Louis Fed On the Economy, Sept. 23, 2024.
- Source needed: World Economic Forum article citing Bloomberg estimates (67% / 21%).
