AI is rapidly reducing the amount of technical skill required to produce many kinds of work.
You don’t need to be a sketch artist to create an infographic. You don’t need to be a web developer to create a clean homepage. And increasingly, you don’t need to be a programmer to create software.
But this does not mean expertise has become irrelevant.
It may mean that expertise and execution are being separated.
1. From Execution to Structural Sense
We have a concept of common sense in everyday life: an intuitive understanding of how things should work, what fits, what doesn’t, and what outcome makes sense.
Different domains may have their own version of this capability. We could call it domain-specific structural sense.
A good writer understands how an argument develops, where emphasis belongs, when prose is repetitive, what a reader needs, and whether an idea is actually compelling.
A designer understands hierarchy, proportion, balance, visual communication, and the relationship between elements.
A programmer understands how software components interact and can recognize when something that appears to work is structurally unsound.
These abilities are not identical to the ability to execute the task manually.
AI increasingly handles execution.
The human increasingly supplies direction, judgment, evaluation, and refinement.
2. My Own Experience with AI
I am not a sketch artist.
I also don’t particularly know how to write sophisticated prompts for image generation. My prompts are often one sentence long.
Yet I can work with AI to produce excellent infographics.
That isn’t because I have no expertise.
The expertise is elsewhere.
I understand the idea I am trying to communicate. I can discuss it in detail, evolve the idea, decide what matters, recognize what is missing, reject a poor representation, and ask for changes.
AI handles the enormous number of possible ways that idea can be turned into a visual artifact.
The same thing happened with our hub website.
I don’t know how to code the website myself. Yet the homepage came out clean because I could reason about what the website needed to accomplish, how information should be organized, what should be prominent, and what felt wrong.
AI translated those requirements into code.
I didn’t need to know how the implementation worked in order to judge whether the result worked.
That distinction is crucial.
3. Vibe Coding and the CS Question
This is why a recent paper examining whether computer-science achievement predicts vibe-coding proficiency is interesting—but also why its interpretation needs to be kept within its actual scope.
Using AI as a code generator is a perfectly legitimate way to study AI-assisted programming.
But programming is only one domain of AI-assisted work.
Coding is not the only job in the world, and software code is not the only medium through which AI produces useful output.
A person’s productivity gain from AI can come from doing something they previously could not do at all.
- Someone who cannot code may nevertheless use AI to create a website.
- Someone who cannot draw may create an infographic.
- Someone without professional design training may create a presentation.
- Someone without programming expertise may automate a workflow.
If we measure only how well people generate code with AI, we are measuring a very specific corner of the much larger AI productivity landscape.
That is useful research.
But it should not automatically become a general theory of AI-assisted human capability.
4. The Problem with the Comparison
If CS graduates outperform ordinary people at AI-assisted coding, that does not necessarily demonstrate that computer-science knowledge itself is responsible.
People who successfully complete demanding CS education may differ from the comparison group in many other ways: academic achievement, analytical habits, general reasoning ability, persistence, familiarity with complex systems, and other cognitive characteristics.
A stronger test would compare CS-trained people with cognitively and educationally comparable people who have substantial expertise in other domains—for example, highly educated professionals, researchers, doctors, physicists, lawyers, or others with advanced training.
Then we could ask a more precise question:
Does specifically computer-science knowledge predict AI-assisted coding performance after broader cognitive and educational differences are controlled?
And even that would answer only the programming question.
It would not tell us whether domain expertise is generally necessary for benefiting from AI.
5. Productivity Is Bigger Than Code Generation
This is where the research question can become misleading if the metric is allowed to stand for something much larger.
Suppose AI makes a person ten times faster at producing code.
That is a significant productivity gain.
But suppose AI enables another person to perform a task they previously couldn’t perform at all.
That may be an even more important transformation.
My own website experience is an example.
The relevant measurement isn’t:
“How much code did I produce?”
I produced essentially none manually.
The relevant measurement is:
“Could I turn my understanding of what I wanted into a useful website?”
The answer is yes.
The AI didn’t merely make me a faster programmer.
It changed the boundary of what I could produce.
That is a fundamentally different kind of productivity gain.
6. AI Does Not Necessarily Make a Bad Writer a Good Writer
The same distinction explains why AI assistance doesn’t automatically eliminate the importance of craft.
AI can make a bad writer’s sentences more fluent.
It can improve grammar, vocabulary, formatting, and surface coherence.
But if the writer doesn’t understand argument, structure, rhythm, audience, emphasis, evidence, or what makes an idea interesting, AI may simply make the underlying weakness more efficiently expressed.
The machine can generate variations.
It cannot guarantee that the person knows which variation is good.
This is where domain-specific structural sense becomes important.
7. What AI Is Actually Changing
The traditional model of expertise looked something like this:
Knowledge → technical execution → artifact
AI increasingly enables a different model:
Domain knowledge + structural sense → direction and judgment → AI execution → artifact
This doesn’t mean technical knowledge disappears.
Sometimes technical knowledge remains essential because the consequences of error are severe or because the task itself depends on deep technical understanding.
But in many domains, the human no longer needs to possess every production skill traditionally associated with the output.
The important capability may move upward.
From:
“Can I make this?”
to:
“Do I understand what needs to be made, and can I recognize when it is good?”
8. The Research Question May Need to Change
This is why I would be cautious about interpreting narrowly designed AI experiments as universal evidence about human capability.
If the research begins with coding, measures coding performance, and finds that CS knowledge predicts coding performance, the result is perfectly compatible with the obvious conclusion:
CS knowledge helps with AI-assisted coding.
But the much larger question is different:
What human capabilities matter when AI becomes the execution layer across many domains?
To answer that, we need experiments involving multiple forms of AI-assisted work:
- writing
- visual communication
- web design
- research
- data analysis
- programming
- presentation design
- problem solving
- creative development
And we need to separate at least three things:
General cognitive ability
Domain-specific structural sense
Medium-specific execution skill
Otherwise, we risk measuring the capability we selected rather than discovering which capabilities actually matter.
9. The Deeper Transformation
Perhaps the biggest mistake is to think that AI is simply making experts faster.
Sometimes it is.
But sometimes AI allows someone to cross a boundary that previously required years of specialized production training.
The artist doesn’t necessarily disappear.
The programmer doesn’t necessarily disappear.
The designer doesn’t necessarily disappear.
Instead, parts of their execution capability become available through an AI transformation layer.
The human can operate at a higher level of abstraction.
That is what I have experienced with images and websites.
I don’t know the implementation details.
But I know enough about the underlying problem to guide the transformation.
And that may be the emerging human skill that matters most:
not knowing how to produce every possible output, but knowing enough about what the output should accomplish to direct, evaluate, and improve what AI produces.
Call it design sense.
Call it domain-specific structural sense.
Perhaps, at a deeper level, it is related to the general cognitive ability psychologists have tried to capture with g.
Whatever we eventually call it, AI is making one thing increasingly visible:
When machines become extraordinarily good at execution, the value of human expertise may shift from making things to knowing what should be made—and recognizing whether what was made is actually good.
10. Education Cannot Remain Fixed
This transformation creates a problem far beyond coding.
Our education systems are largely built around relatively stable assumptions about what people need to know. A curriculum is designed, a course lasts several years, a qualification is awarded, and society assumes that the acquired skills will remain valuable for a reasonable period of time.
AI challenges that assumption.
What is necessary today may become partially redundant tomorrow. What is difficult today may become trivial tomorrow. And entirely new capabilities may become important that did not previously justify a place in the curriculum.
This is increasingly being recognized within MIT itself. MIT’s leadership has called for an AI-aware educational process that revisits what students need to learn and for systems of continuous reflection, iteration, and improvement so education can adapt to rapid AI change.
MIT: AI and education — A watershed moment for MIT (official link): https://president.mit.edu/writing-speeches/ai-and-education-watershed-moment-mit
MIT Open Learning has similarly emphasized the importance of developing adaptable, AI-fluent professionals rather than relying solely on narrow skills that may quickly become obsolete.
MIT Open Learning: Creating adaptable, AI-fluent professionals (official link): https://openlearning.mit.edu/news/new-online-learning-experience-aims-create-adaptable-ai-fluent-professionals
The question should not simply be:
“What should a student learn?”
It should increasingly be:
“What should a student learn given what AI can already do—and what AI is likely to do next?”
That does not mean teaching less.
It may mean teaching different layers of knowledge.
Students may need less time mastering routine execution that AI can reliably perform and more time developing conceptual understanding, structural sense, critical evaluation, problem formulation, creativity, judgment, and the ability to work with AI as a transformation layer.
A four-year curriculum designed today should not automatically remain the correct four-year curriculum simply because it has always been four years.
Education should continuously reassess what is essential, what is becoming redundant, and what new capabilities have become necessary.
The future may therefore require a shift from fixed education to adaptive learning.
Not a system in which people learn once and then rely on a qualification for decades, but one in which education continuously responds to the changing boundary between what humans need to know and what machines can do.
11. From the Small Boat to the AI Ship
Perhaps the simplest way to understand this transition is through an analogy.
For much of human history, we were like people in a small boat on a relatively calm lake.
We had limited tools and limited access to knowledge. Our primary responsibility was to operate the boat ourselves: rowing, adjusting the sail, watching the water, and making almost every operational decision ourselves.
Education therefore had a natural emphasis on learning those skills.
AI changes the vessel.
We are increasingly moving into an advanced AI-powered ship on a vast sea of knowledge.
The ship can perform tasks that previously required enormous amounts of human effort. It can generate, transform, search, analyze, design, calculate, and navigate through possibilities at a scale no individual human could manage manually.
In that environment, our role shifts from continuous manual control to equally important oversight.
It is similar to an advanced aircraft with autopilot. The pilot may no longer be continuously manipulating every control surface, but that does not make the pilot irrelevant. The pilot still has to set direction, monitor the systems, recognize anomalies, intervene when necessary, and decide where the aircraft should go.
The nature of the responsibility has changed—from manual operation to supervision, judgment, and intervention.
AI therefore doesn’t necessarily eliminate human responsibility.
It can move responsibility one level upward.
The ship will keep becoming more capable.
The sea of knowledge will keep becoming larger.
And the skills required to operate the ship will keep changing.
That is why education cannot remain fixed.
We should not keep training humans simply to row faster after we have built a ship. We should teach them how to understand, supervise, maintain, and intelligently direct the ship.
And this brings us back to the CS degree.
If AI continues to improve at generating, debugging, and explaining code, will the productivity advantage gained from spending four years learning computer science remain large enough in the future to justify those four years specifically for coding?
Or will the value of those four years increasingly come from something deeper—the structural understanding of computing that remains useful even after AI has taken over much of the code production?
And if the answer changes again five years later, should education itself be designed to change with it?