For most of human history, expertise was scarce. A person spent decades learning mathematics, medicine, law, engineering, teaching, writing or research. That accumulated knowledge became valuable because another person had to acquire it slowly. Society therefore developed an elaborate system for transferring expertise through schools, universities, professions, examinations and, ultimately, the expert’s time.
AI introduces a strange possibility. The knowledge accumulated by one human may increasingly be reproduced by a machine that can make it available to millions of people simultaneously.
The consequence is not simply that AI makes experts more productive. AI may change the economics of expertise itself.
1. The Moment the Teacher Stops Being the Expert
For centuries, education rested on a simple asymmetry: the teacher knows more than the student. A good teacher accumulated knowledge, experience and teaching skill over decades, while the student paid for access to that accumulated expertise. The model worked because expertise was scarce.
AI is beginning to challenge that scarcity.
A September 2026 StudentBench study compared AI tutoring with expert human tutoring across 2,383 participants preparing for the GRE. The researchers found statistically equivalent learning gains between pooled AI and expert human tutoring in the study, while the best-performing AI tutor exceeded the human tutor’s average result in five of seven GRE domains. One AI tutor achieved equivalent measured gains at a reported 918× lower cost per percentage point gained: $0.0052 versus $4.81.
The study is narrow. It concerns GRE preparation, short tutoring interventions and immediate learning gains. It does not demonstrate that AI can replace teachers in every educational setting. But it exposes a much larger question: what happens when the scarce expertise that justified the teacher’s price can be reproduced by a machine at almost negligible marginal cost?
The issue is no longer simply whether AI can help the teacher. It is whether the teacher is still required for the particular function being purchased.
2. AI Is No Longer Just an Assistant
The usual description of AI is comforting: AI will assist humans. The doctor will use AI, the lawyer will use AI, the programmer will use AI, the teacher will use AI and the researcher will use AI.
There is nothing wrong with this description. It may accurately describe the first stage of the transition. But it hides an important possibility.
An assistant is useful because the human remains the principal source of the valuable function. What happens when the assistant becomes capable of performing that function itself?
A calculator did not merely make arithmetic faster; it removed most of the need to perform arithmetic manually. A spreadsheet did not merely help accountants; it changed what accounting work meant. Search engines did not merely help people remember facts; they reduced the economic value of memorizing enormous quantities of information.
AI extends this pattern into cognitive work. The distinction is therefore becoming one between AI assisting the expert and AI performing the expertise. Those are not the same technological transition. This connects with the broader question explored in what becomes valuable when knowledge becomes abundant.
3. The Tutoring Shock: When AI Matches Expert Humans
The tutoring experiment is interesting because it measures something more meaningful than whether an AI can answer a mathematics question. It asks whether the student actually learns.
That distinction matters. An AI can produce an impressive explanation and still be a poor teacher. It can solve a problem correctly while failing to identify the student’s misconception. It can generate sophisticated material that teaches almost nothing.
StudentBench attempted to measure the outcome on real students rather than simply judging AI responses. Across the study, AI tutoring and expert human tutoring produced statistically equivalent GRE learning gains. The experiment involved 2,383 participants and thousands of tutoring sessions.
There are important qualifications. The experiment does not establish equivalence across all ages, subjects, educational environments or long-term outcomes. The researchers frame the result around the particular GRE tasks studied.
But that is precisely why the result is useful. It does not prove that teachers are obsolete. It demonstrates that one of the things we thought required an expert human—individualized tutoring—can already be performed by AI at a comparable level under at least one controlled condition.
That is enough to change the question.
4. 918× Cheaper: The Economics of AI Tutoring
Capability is only half of the story. A human tutor may be extremely good, but a human tutor has a cost structure that cannot be escaped. One teacher teaches one student, or a small group, for a finite number of hours.
AI has a fundamentally different cost curve.
The StudentBench paper reports one AI tutor achieving statistically equivalent learning gains to expert human tutoring at $0.0052 per percentage point of gain, compared with $4.81 for human tutoring, using the study’s $75-per-hour human-tutor reference. The authors describe this as a 918× cost difference.
That number should not be interpreted as a universal market price for AI tutoring. It is a study-specific cost reconstruction based on the experimental setup and pricing assumptions. Nevertheless, the economic principle is straightforward.
If two systems produce comparable outcomes and one requires dramatically less human labour, the expensive system cannot simply defend its price by saying, “But I am the expert.”
The market eventually asks a different question: “What additional outcome am I buying?”
That is a much harder question for traditional expertise.
5. When AI Beats the Teacher
There is an even more uncomfortable result in the StudentBench data. The study reports that in five of seven GRE domains, the best-performing AI tutor surpassed the human tutor’s average learning gain.
Again, this does not mean that AI tutors are universally better than human teachers. It means that under the measured conditions, the strongest AI system tested performed better than the human benchmark in those domains.
But imagine the direction of travel.
Suppose a student can ask an AI the same question ten different ways. The AI can identify where the student’s reasoning went wrong, generate a new problem at exactly the right difficulty and immediately test whether the student has understood. It can repeat the process indefinitely, without the student having to schedule another lesson.
At that point, the human teacher is no longer competing simply against another teacher. The teacher is competing against a system that can potentially replicate parts of the teacher’s expertise at machine scale.
6. From Expert Tutors to Expert AI
This creates a subtle transformation in the meaning of expertise.
Consider a mathematics teacher with twenty years of experience. That experience contains thousands of observations: which mistakes students repeatedly make, which explanations work, which shortcuts create misconceptions, which examples make an abstract idea intuitive and which sequence of problems takes a student from confusion to mastery.
Traditionally, all of that knowledge remained attached to the person.
AI changes the possibility. The teacher’s expertise can become input to a system rather than the final product delivered directly by the teacher. The teacher may still be the origin of valuable knowledge, but the machine can become its distribution mechanism.
One expert teaches a hundred students. Then perhaps a thousand. Then potentially millions.
The economic relationship changes from one expert serving many students to one expert’s knowledge flowing through AI to millions of students.
The scarce expert has not necessarily disappeared. The scarcity of the expert’s time has.
7. The Mensa Problem: What Happens When AI Beats Very High-IQ Humans?
Tutoring is only one example. A deeper question arises when AI begins to perform well on tests designed to measure human reasoning itself.
IQ-style comparisons are imperfect. They compress a complicated collection of cognitive abilities into a single number, and mapping an AI benchmark score onto a human IQ distribution is not scientifically equivalent to measuring the IQ of a person.
Nevertheless, researchers have begun analysing LLM performance on IQ-style evaluations, including the Mensa Norway test. Some recent analyses report that advanced models perform within very high portions of the human distribution on such evaluations. Those comparisons should be treated as benchmark comparisons rather than literal measurements of an AI’s human IQ.
The useful question is not, “What is the AI’s IQ?” The more interesting question is: what happens when a machine can perform cognitive tasks that previously required unusually high human reasoning ability?
Imagine that a company once needed a small number of exceptionally capable people to solve a particular class of problems. If AI can eventually perform those problems reliably, the organization does not need to make every employee exceptionally capable at that particular task. It can give ordinary people access to extraordinary cognitive tools.
AI may therefore simultaneously reduce the economic scarcity of intelligence while increasing the importance of deciding what that intelligence should be used for.
8. Intelligence Has Become Abundant
Human civilization has always been constrained by the supply of capable minds. There are only so many brilliant mathematicians, outstanding teachers, experienced engineers and researchers who can spend thousands of hours becoming exceptional at a narrow field.
AI potentially changes that supply curve. The idea also connects with AI as an intelligence multiplier: the important shift is not simply better individual performance, but the multiplication of cognitive capability.
A capable model can be copied. It can operate continuously. It can serve millions of users. It can be upgraded without every user having to acquire another twenty years of experience.
This creates a new economic possibility in which intelligence becomes abundant.
But intelligence becoming abundant does not mean everything becomes abundant. The ability to solve a problem is not the same as deciding which problem is worth solving.
A machine may be able to produce a thousand business ideas, but someone still has to decide which business to build. AI may generate ten thousand research hypotheses, but someone still has to decide which question deserves a decade of attention. AI may produce a perfect explanation, but someone still has to decide whether the student should learn it.
The scarcity may therefore migrate.
9. AI Will Augment Everything
Your original formulation is stronger than the usual claim that AI will affect many jobs: AI will augment everything.
Almost every cognitive activity can potentially acquire an AI layer. Writing gets AI. Programming gets AI. Teaching gets AI. Research gets AI. Design gets AI. Analysis gets AI. Planning gets AI. Management gets AI.
At first, this looks like universal productivity enhancement, and in many cases it will be exactly that.
But there is a hidden asymmetry. Suppose a human becomes twice as productive because of AI. That is augmentation. Now suppose another generation of AI makes the system ten times more capable and performs most of the task itself.
The human’s productivity may rise again, while the human’s share of the work falls.
This distinction is crucial: human output increasing does not necessarily imply human contribution increasing. A person can produce more while doing less of the underlying work. That is where augmentation starts becoming something else.
10. But Augmentation Can Become Substitution
Augmentation and substitution are not opposites. They can be consecutive stages.
A calculator augmented the mathematician and then substituted for human arithmetic. Software augmented the accountant and then substituted for much manual bookkeeping. Search engines augmented memory and then substituted for much of the need to memorize information.
AI may follow the same pattern at a much larger scale.
The sequence can look like this: AI helps the human perform the task; AI performs most of the task while the human supervises; the human checks the AI’s work; eventually, the human may only decide whether the task should be performed at all.
That is why “AI will augment workers” is not necessarily reassuring. Augmentation can be the road to substitution.
The real question is not whether AI augments the worker. It is how much of the original human function remains after the augmentation is complete.
11. The Human-to-AI Transition
The transition can be represented simply: Human → Human + AI → AI-supervised human → AI + human oversight → AI.
Not every occupation will travel through every stage. Some tasks will stop at human-plus-AI because human interaction remains essential. Others may move much further.
Consider software development. The programmer once wrote almost every line. Then came libraries, autocomplete, code generation and increasingly capable AI agents that can modify multiple files, run tests and debug their own work.
The programmer has not necessarily disappeared. But the unit of human contribution has changed.
The same pattern can occur in education. The teacher moves from explaining every concept to supervising an AI tutor. The human is still present, but presence and necessity are different things. For the deeper question of where AI gets its answers, see Where Does AI Look for the Answer?.
12. If Knowledge Is No Longer Scarce, What Is the Value of Knowing?
For most of history, knowing something was valuable because acquiring that knowledge was difficult.
A person who knew calculus was unusual. A person who knew several languages was unusual. A person who had mastered a complicated profession was unusual. Education was therefore partly a process of transferring scarce information and methods from one generation to another.
AI attacks that scarcity.
If a student can ask an AI to explain calculus at beginner, intermediate or advanced level, the knowledge itself becomes easier to access. If a programmer can ask an AI how to implement an unfamiliar algorithm, memorizing the implementation becomes less valuable. If a researcher can ask an AI to summarize thousands of papers, the ability to manually search and summarize literature becomes less scarce.
Knowledge does not become worthless. Its scarcity value changes.
Knowing something may still matter enormously, but the economic advantage of being the only person who knows it becomes weaker.
That forces a difficult question: if everyone can access the answer, what exactly does expertise mean? It is closely related to the problem explored in The Limits of AI: When Knowing the Answer Is Not Understanding.
13. What Humans Still Control: Goals, Purpose, Judgment and Agency
The answer may lie in a distinction that AI capability alone cannot erase: solving a problem is different from choosing the problem.
A chess engine can calculate positions, but it does not thereby decide whether humanity should play chess. An AI can design a pharmaceutical compound, but it does not thereby decide which disease society should prioritize. An AI can write a book, but it does not establish why the book should exist.
These distinctions should not be romanticized. AI systems increasingly participate in planning, evaluation and decision-making, and the boundary between execution and direction is becoming less clear.
But direction remains conceptually different from execution. Someone has to establish the objective, accept responsibility for the consequence and decide what counts as success.
This creates a possible new hierarchy: AI supplies capability; humans supply objectives.
Whether humans retain that role in practice is not guaranteed. But if intelligence becomes abundant, agency may become more valuable than intelligence itself.
14. The Collapse of the Traditional Expertise Premium
For centuries, society paid people partly for what they knew. Doctors knew medicine. Lawyers knew law. Teachers knew their subjects. Engineers knew engineering. Consultants knew how to solve particular problems.
The premium came from accumulated expertise.
AI introduces another possibility: expertise becomes embedded in a system that anyone can access.
The consequence may not be the disappearance of experts. Instead, the premium may split. Some forms of expertise could become cheap and abundant, while other forms may become more valuable precisely because they are difficult to automate.
The expert who merely supplies information becomes vulnerable. The expert who understands context, takes responsibility, establishes trust, makes difficult judgments and knows what not to do may remain valuable.
This produces a strange inversion. The more capable AI becomes at answering questions, the less valuable it may be to possess answers. The more capable AI becomes at execution, the more valuable it may become to know which execution is worth paying for.
15. What Happens to Education When the Teacher Is No Longer the Best Teacher?
Education may be one of the clearest places where this transition becomes visible.
The traditional classroom is built around scarce teaching capacity. Thirty students share one teacher. One student raises a question while twenty-nine students wait. The teacher has to choose an explanation that works for the average student.
AI can potentially reverse that structure.
Every student can have a private tutor. The explanation can change from student to student. The pace can change. The examples can change. The system can remember previous mistakes, and the student can ask the same question repeatedly without embarrassment.
The teacher can potentially see the entire learning process rather than only the student’s final answer.
That does not make teachers irrelevant. It changes the job. A teacher might increasingly spend less time transferring information and more time creating an environment in which learning happens.
But there is an uncomfortable possibility. If AI eventually becomes better at explanation, diagnosis and adaptive practice, the teacher may cease to be the best source of instruction while remaining an important part of education.
Those are very different propositions.
16. What Happens to Work When the Expert Is No Longer the Bottleneck?
The same logic extends beyond education.
Imagine a small company with ten people. In the past, it might have needed a finance specialist, a designer, a programmer, a researcher, a copywriter and several managers. Now imagine those ten people working with AI systems capable of performing large portions of each function.
The company does not simply have ten employees who are more productive. It may have ten people with access to the functional capacity of a much larger organization.
This is one reason AI could produce an unusual organizational phenomenon: a small human team plus large AI capability can generate large economic output.
The bottleneck moves.
When expertise is scarce, organizations need experts. When expertise is abundant, organizations need fewer people to access it.
The limiting factor becomes increasingly likely to be capital, distribution, physical resources, customers, regulation, trust and human decision-making rather than the number of people who can perform routine cognitive tasks.
That could make the smallest organizations disproportionately powerful. The company of the future may be measured less by how many people it employs and more by how much economic output its human-AI system can produce.
17. The New Scarcity: Direction Rather Than Knowledge
Imagine giving a person access to an AI capable of answering almost any question. What becomes scarce?
Not answers. Not drafts. Not calculations. Not explanations. Perhaps not even ideas.
The scarce resource may become direction.
A person can ask AI, “What should I do?” and receive ten plausible answers. But choosing among them remains a human problem.
The next question becomes, “Why do I want this outcome?” That question is harder.
AI can optimize a goal, but goals themselves have consequences. A company can optimize profit. A researcher can optimize publication output. A student can optimize examination scores. A government can optimize some measurable social indicator.
Yet optimization without judgment can produce absurd results.
The future may therefore reward people who can define meaningful objectives before asking machines to optimize them. The scarce skill becomes less “Can you solve this?” and more “Should this be solved—and if so, why?”
18. From Human-Centred Organizations to AI-Centred Organizations
The industrial revolution reorganized production around machines. Factories were not simply offices with better tools; the machine became the organizing principle of production.
AI could produce an analogous change in knowledge work.
Today’s organization is still fundamentally human-centred. Humans occupy roles, humans perform functions and software supports them.
AI may reverse that arrangement.
The organization could increasingly be built around AI systems, with humans occupying a smaller number of strategic positions around them. Instead of 100 employees using software, we may see 10 employees directing AI systems that perform the work of a much larger organization.
This does not mean every company will become tiny. Some activities require physical infrastructure, regulation, sales networks or large human communities.
But where cognitive work dominates, the relationship between workforce size and output could change dramatically.
The company of the future may be measured less by how many people it employs and more by how much economic output its human-AI system can produce.
19. The Economic Problem After Expertise Becomes Cheap
This is where the story becomes larger than technology.
If AI makes expertise dramatically cheaper, society gets an enormous productivity opportunity. But productivity and employment are not the same thing.
A machine that makes one worker ten times more productive does not automatically create ten times as many jobs. Historically, technological progress often created new industries and occupations, but AI is unusual because it is aimed directly at many of the cognitive functions through which humans traditionally adapt to technological change.
The central question becomes: what happens to people when their economically valuable expertise is no longer scarce?
That question cannot be answered simply by saying that everyone should “learn AI.” If AI eventually augments nearly every profession, learning to use AI may itself become a temporary advantage rather than a permanent occupation.
The deeper issue is the distribution of the productivity created by AI. Who owns the systems? Who owns the capital? Who receives the income? Who decides what the machines produce?
And perhaps most importantly: what does human work mean when work is no longer required for much of what humans are capable of doing?
That is the economic problem hiding underneath the tutoring example.
20. The Bigger Question: What Is Human Value in an Age of Superhuman AI?
We began with a teacher. A teacher knew more than the student. Then AI learned enough to teach the student. Then, in some measured tasks, AI approached or exceeded the performance of the human benchmark.
The same pattern may appear across other forms of expertise.
That does not prove that humans are becoming useless. It points to something more subtle and potentially more consequential: the historical relationship between intelligence and economic value is changing.
For thousands of years, human intelligence was scarce. Education increased it. Experience refined it. Institutions organized it. Organizations paid for it.
Now machines can reproduce increasing amounts of cognitive capability at extraordinary speed and potentially enormous scale.
The consequence may be a world in which intelligence is no longer the primary scarce resource.
If that happens, humanity faces a strange inversion. We spent civilization trying to become more intelligent. We may now have to learn what to do after intelligence becomes abundant.
The teacher may no longer be the person who knows the answer. The programmer may no longer be the person who writes the code. The researcher may no longer be the person who performs every analysis. The expert may no longer be the person who possesses scarce knowledge.
Their value may increasingly lie somewhere else: choosing the goal, understanding the context, exercising judgment, accepting responsibility, creating meaning—and deciding what the extraordinary new machinery of intelligence should actually be used for.
That is the real question behind AI beating the expert.
Not “Can the machine replace the human?” but “When the machine can do what made the human valuable, what will make the human valuable next?”
Editorial note: The tutoring figures discussed in this article refer to the 2026 StudentBench study and its reported experimental cost reconstruction. The study concerns a specific GRE tutoring intervention and should not be interpreted as proof that AI is universally superior to human teachers or that all forms of education can be automated.
Read Next
Is AI Ending the Moat of Large Organizations?
What happens when AI makes a small human team capable of doing work that once required a much larger organization?