For babies being born in recent years, AI will not arrive as a sudden technological disruption. By the time they begin forming durable memories, systems that can converse, explain, generate, and offer advice will already be woven into everyday life, as ordinary as search, maps, and video.
That raises a question about education that feels more fundamental than whether students should be allowed to use AI, how homework should change, or which tools schools ought to teach:
If external intelligence is available from the very beginning, how does a person form a mind of their own?
Many human capacities are not built by simply acquiring correct information. A young child touches things, falls down, and tries again. A student writes a first draft they do not particularly like and begins to understand the problem by revising it. Someone early in a career spends years doing work that may appear mundane, gradually learning where things tend to go wrong and which details matter.
These processes are inefficient. They involve confusion, repetition, failure, and long periods of being stuck. Yet judgment, patience, problem sensitivity, experience, and self-efficacy often emerge precisely through those moments. The deepest effect of AI on education may therefore be not merely that it accelerates access to knowledge or the production of content. It is that AI is beginning to intervene in the pathway through which capability itself is formed.
Designing the Right Friction
Traditional learning contains plenty of low-value friction. Searching through ten irrelevant pages to find a single useful source, spending hours fixing formatting, or mechanically repeating something already understood dozens of times are not experiences we need to preserve. AI can compress much of this work, and that is a good thing.
But other forms of friction serve a different purpose. The pause when a problem has not yet made sense. The uncertainty of encountering an open-ended question for the first time. Making a choice, seeing that the outcome is not what you expected, and revising your judgment afterward. These experiences are not merely obstacles to learning. They are part of learning.
The question has never been whether more difficulty is inherently better. It is whether a particular difficulty is simply consuming time or actively shaping the way someone thinks.
While discussing this with AI, it gave me a phrase that stayed with me:
Don’t remove all friction. Design the right friction.
Education does not need to preserve old forms of pain simply to prove that humans are still thinking. Nor does it need to treat AI as a contaminant that must be isolated from the learning process. The real design problem is to decide where AI should reduce unnecessary work and where learners still need to encounter choice, ambiguity, feedback, and consequences.
If learning is a journey, AI can remove many pointless detours. The responsibility of education is to make sure the road still contains moments that require observation, judgment, course correction, and commitment.
Before Explanation Comes Experience
This question becomes especially important in early childhood.
A child does not build a model of the world by beginning with explanations. A leaf is not first understood as information about airflow. It is something that moves, falls, can be touched, and responds when you blow on it. Water is not initially H₂O. It is wet. It flows. It spills out of a cup.
Human relationships are not first encountered as a system of social rules either. They are experienced through sharing, refusal, waiting, conflict, separation, and repair.
Much of our most fundamental knowledge begins as embodied and relational experience before it ever becomes language.
A child growing up with AI may be able to ask “why?” and immediately receive a complete, age-appropriate, beautifully articulated answer. The answer may be perfectly correct. The question is whether it arrived at the right moment.
If a child asks why leaves move, we could immediately have AI explain wind. Or we could first look at different trees, blow on a leaf, notice that a stone does not move in the same way, and let another question emerge. AI can enter afterward—not to close the inquiry, but to extend it. It might ask what else the child could try, connect the observation to something else they care about, or help a parent think of another way to investigate the phenomenon.
For young children, perhaps the most useful role for AI is not an always-available voice explaining the world, but a curiosity extender.
What needs protecting at this stage is direct experience: the relationship between the child and objects, bodies, nature, other people, and real consequences. A child needs to develop a very basic sense that the world is available to be observed; that I can try something myself; that I can make a wrong guess; that I can also discover something.
That feeling later becomes an important foundation for agency.
One way to describe the process is:
Experience → Wonder → Ask → Try
AI can appear at any point in that sequence. But ideally, it helps the child return to the world rather than allowing language to replace the world.
When the Output No Longer Proves the Learning
The problem changes when children become students.
Essays can now be generated. Research can be generated. Even a polished simulation of a “thinking process” can be generated. In the past, we often treated the quality of an artifact as a proxy for the capability behind it. A strong essay usually implied that a student had done at least some reading, organization, and reasoning. A complete project usually suggested that they had gone through some meaningful process of making.
That connection is no longer reliable.
But the answer should not be to turn every assessment into an oral exam, an in-class performance, or an exercise in proving that the work was done “by yourself.” Education can very quickly collapse into an anti-cheating system if authenticity becomes its primary design objective.
The more important question is this:
What kinds of problems and processes allow AI to participate fully while still requiring learners to perform meaningful cognitive work?
If an assignment is simply “produce X,” whether X is an essay, presentation, or product proposal, AI can take over much of the path between prompt and output. More valuable learning environments tend to contain real tradeoffs and real friction. Goals conflict. Information is incomplete. New evidence disrupts earlier assumptions. Different people disagree about what a good outcome even means.
AI can analyze, generate, simulate, and recommend. The learner still has to decide which problem matters, which evidence deserves trust, which tradeoffs are acceptable, what new information actually changes, and which direction they are willing to stand behind.
A useful sequence might be:
Frame → Encounter → Reconsider → Commit
Frame does not mean requiring students to complete the problem once without AI. It means entering the problem with some initial orientation.
Encounter means allowing that initial understanding to collide with something real: an alternative interpretation generated by AI, peer criticism, new evidence, user feedback, or a suddenly changing constraint.
Reconsider does not mean performing the ritual of “changing one’s mind.” It means judging what should change and what still holds.
Commit means eventually arriving at a position, artifact, decision, or action the learner is willing to own.
The final output may involve extensive use of AI. What matters is that the learner still carries responsibility for the choices within it.
Education as Decision Construction
Seen this way, the unit of educational design is no longer simply the assignment or the test. It begins to look more like the construction of decisions.
A good curriculum should not exhaust itself trying to find questions AI cannot answer. That boundary will continue moving. A more durable anchor is to ask:
What must this person ultimately become capable of doing?
If the goal is judgment, then the curriculum must create the conditions under which judgment is exercised. If the goal is problem definition, students must encounter problems that have not already been neatly carved into solvable pieces.
If the goal is understanding complex systems, students cannot simply be asked to submit answers that look complete.
This leads naturally to the question of grounding.
AI can generate a remarkably convincing set of user insights in seconds. Talk to five actual users, and the neatness begins to unravel. A business plan can appear perfectly coherent inside a model while real stakeholders remain completely unconvinced. A design can satisfy every best practice and still encounter resistance once it reaches a real environment.
These moments are not noise in the educational process. They are learning material. They force students to reconcile abstract knowledge with the world as it actually behaves.
AI has access to vast amounts of modeled knowledge about the world. Humans develop situated understanding through particular experiences. Good education should repeatedly bring the two into contact.
“Real-world projects” should not be romanticized either. Simply throwing students into reality does not guarantee growth. Reality itself needs to be designed as part of the learning environment. When should users enter the process? When should a new constraint appear? When should students be forced to revisit an earlier assumption? These are learning-design decisions.
The important thing is that learning should not remain entirely inside a closed textual environment that can be cleanly simulated. The irregular, unpredictable, shifting qualities of reality are themselves part of how judgment develops.
What Happens When AI Removes the Junior Work?
Eventually, students enter professional life, and here education and organizations face another problem: AI is rewriting traditional apprenticeship.
A junior designer once spent years doing production work. An analyst spent long hours cleaning and organizing data. A young strategist collected references, assembled decks, and revised them again and again. Much of this work was low-value labor. But it also performed another function: it was how people accumulated experience and learned to recognize patterns.
Do enough work and you begin to know where mistakes tend to occur. Watch a senior revise enough documents and you begin to see why something can look perfectly reasonable yet still be useless. Work on enough real projects and you begin to learn which rules stop working under which conditions.
These are also precisely the kinds of tasks AI tends to automate first.
Organizations therefore cannot ask only, “How do we make juniors do less of this?” They also need to ask: Where will the capability-building function previously hidden inside this work now take place?
The answer is not to preserve every old junior task. Quite the opposite. AI should remove much of the repetitive work, while the underlying training objective becomes more explicit.
If a junior was once asked to find twenty examples, the actual skill being developed may not have been search itself, but relevance judgment. AI can now find fifty examples. The junior can instead decide which ones should be excluded, what important evidence is missing, and what the material actually demonstrates.
If writing repeated drafts once helped develop a sense of structure, AI can now produce several versions instantly. The junior can compare them, diagnose why one structure works and another is merely superficially coherent, and rebuild the argument.
The tacit knowledge of senior practitioners also needs to become easier to observe. Experts often look at something and immediately feel that “something is off.” What a beginner needs is access to what that glance is actually detecting.
The learning loop can therefore shift toward:
Observe → Diagnose → Decide → Reflect
Watch how experts make judgments. Diagnose AI output. Gradually take responsibility for real decisions. Then compare your reasoning with what the expert saw.
AI can allow junior practitioners to enter more complex work earlier—but only if the organization is deliberately designing for judgment formation rather than assuming that because the deliverable looks sophisticated, the person producing it has become sophisticated too.
For Experts, AI Should Be a Challenger
The problem looks different again for people who have been working for ten or twenty years.
Experienced professionals possess a great deal of situated knowledge that AI cannot easily acquire on its own. They know where a client will actually resist. They know which project risks will never appear in a proposal. They know why a perfectly standard solution simply will not work inside a particular organization.
That knowledge is valuable. It also creates grooves.
Experience teaches us where to look first. It makes certain variables more salient than others. It encourages us to carry successful methods from one situation into the next.
At this stage, AI should not function merely as a faster assistant. It can help practitioners make implicit assumptions visible, construct counter-scenarios quickly, introduce developments from adjacent fields, and force someone to articulate why they believe what they believe.
The purpose of learning at this stage is to keep expertise fluid: not to discard accumulated experience, but to continuously update it.
For mature professionals, AI is most valuable as a challenger—not as a replacement for expertise, but as a way of keeping expertise dynamic.
There Is No Single AI Literacy
Put these stages together and it becomes difficult to imagine “AI education” as a single AI literacy curriculum.
Different stages of life require different forms of human development.
Young children need to build a sense of self and a model of the world. Students need to develop reasoning and judgment. People entering professional life need to build expertise. Mature practitioners need to keep existing expertise adaptive.
AI’s role should change accordingly. It should not remain the same generic “intelligent assistant” from early childhood through adulthood.
These four stages are ultimately variations of the same question:
What does a person still need to experience, in the presence of AI, in order to continue growing?
Education eventually has to return to the question of what kind of person we are trying to develop.
Can someone observe the world closely? Can they form their own questions? Can they operate in situations without standard answers? Can they revise their judgment when new information appears? And when vast amounts of external intelligence are available, can they still take responsibility for a decision?
The most important question for education in the age of AI may therefore not be what humans can do that AI cannot.
It is:
What must humans become capable of doing?
Only after that do we design the questions, the processes, the experiences, and the friction.
AI can participate throughout. It can explain, generate, challenge, simulate, give feedback, and help someone reflect on what they have done. What education has to design is the meaningful work that remains the learner’s responsibility inside that process.
If we get this right, AI will not simply compress learning into faster production. It can become a genuine amplifier of human capability. A child can notice more because of AI. A student can encounter more perspectives. A beginner can enter complex work earlier. An expert can be continually confronted with new possibilities.
The crucial question is whether the growth of external intelligence is accompanied by the growth of human agency, judgment, and expertise.
Education in the age of AI should not be about teaching people to complete tasks alongside AI.
It should be about helping people enter a more complex world with AI—and still remain capable of forming their own questions, their own judgments, and their own actions.
About the author
Qinqin (Stella) Yang is a creative AI builder and experiential creative director. She builds workflows, systems, and creative direction that operate under real constraints, and designs speculative frameworks that expand what teams can imagine before they start making. Over the past decade she has built brand experiences, product design, and speculative projects for Tencent, IKEA SPACE10, Burberry, and the Palace Museum. She is the founder of Actual Fiction, where she develops AI-native tools that turn imagination into a working method.
This newsletter is where she thinks out loud about creativity, AI, and what it means to build toward a future that doesn’t exist yet.
More at qinqinyang.com · Practice at Actual Fiction