The Wrong Question About AI in Schools

Aug 28

The debate is not whether students will use artificial intelligence. It is which parts of thinking schools should refuse to make optional.

For a while, the most urgent question about artificial intelligence in schools seemed obvious:

Should students be allowed to use it?

Some schools blocked it. Others embraced it. Teachers rewrote policies. Students learned very quickly that a blocked website on a school laptop was still available on the phone in their pocket.

The argument became familiar.

AI is cheating.

AI is the future.

Students need to learn without it.

Students need to learn how to use it.

All of these positions contain something reasonable.

I think they also miss the more interesting question.

Because once a technology becomes capable of writing an essay, explaining a theorem, generating an argument, translating a passage, brainstorming ideas and correcting mistakes, asking whether students should "use AI" becomes almost meaningless.

Use it for what?

There is an enormous difference between a student asking:

Write my conclusion.

and:

I think my conclusion is weak. Ask me three questions that might expose what I haven't considered.

Both students used artificial intelligence.

Only one may have surrendered the intellectual work the assignment was designed to create.

So perhaps the question schools should be asking isn't:

Should students use AI?

It is:

What work must the student still do?

That distinction could change education far more than any ban.

The essay was never really the point

Imagine a teacher assigns a 1,500-word essay.

For decades, the finished paper served two purposes at once.

It was the student's product.

But it was also evidence.

If a student produced a thoughtful argument, used evidence well, anticipated objections and wrote clearly, the teacher could reasonably infer that something had happened inside the student's mind.

The essay was a kind of intellectual footprint.

Not perfect evidence, of course. Parents helped. Friends edited. Tutors intervened. Some students plagiarized.

Still, the connection between producing good work and being capable of producing good work was reasonably strong.

Generative AI weakens that connection.

A beautiful essay is no longer particularly strong evidence that the person whose name appears at the top can write a beautiful essay.

That creates a much bigger problem than cheating.

It creates an evidence problem.

Schools have spent generations building assessment systems around visible outputs: essays, reports, homework, presentations, code, problem sets.

AI can increasingly produce the output without reproducing the learning that was supposed to precede it.

Which means we may have to separate two things education has historically treated as almost identical:

the quality of the work

and

the capability of the student.

The future of assessment may depend on learning how to measure the second.

We may need to move from proof of production to proof of learning

This could make school surprisingly more human.

Suppose a student submits an excellent history essay.

Instead of asking whether AI touched it, the teacher spends five minutes talking to the student.

Why did you choose this argument?

Which source changed your mind?

What's the strongest objection to your position?

If I removed this paragraph, what would happen to the argument?

What did the AI suggest that you rejected?

Suddenly, authorship becomes much harder to fake.

But something more important happens.

The conversation reveals whether the student can think with the material.

UNESCO has already argued that generative AI forces educators to reconsider what is worth assessing, particularly when machines can reproduce many of the outputs traditionally used as proxies for knowledge. Its work points toward greater emphasis on higher-order reasoning, creativity and judgment rather than simply evaluating finished products.

That could mean more oral defenses.

The problem isn't that AI makes things easy

It is tempting to describe this as a battle against convenience.

But easy is not automatically bad.

Nobody wants students wasting hours looking for a definition an AI tutor can explain in thirty seconds.

A teenager who doesn't understand a physics concept may benefit enormously from an infinitely patient tutor that can try another analogy without embarrassing the student in front of the class.

There is real promise here.

A 2026 randomized controlled trial involving 371 students in Grades 7 to 9 tested generative-AI support during physics and English classes. Researchers found that certain carefully designed prompts helped preserve students' sense that the material was useful and personally relevant. At the same time, the interventions did not produce clear improvements across several other outcomes, including domain knowledge and some learning strategies.

That result is less dramatic than either side of the AI debate might like.

AI did not destroy learning.

It did not magically transform it either.

What mattered was how the system was designed to interact with the student.

And that may be the lesson.

The goal should not be to maximize the amount of AI in education.

Nor should it be to minimize it.

The goal should be to decide where convenience supports learning—and where inconvenience is the learning.

Schools may need to make AI less helpful

That sounds ridiculous.

We have spent decades designing technology to be more useful.

But one of the most interesting educational ideas I have encountered recently goes in exactly the opposite direction.

A small 2026 study of high school writing deliberately constrained AI chatbots so that they would ask students questions instead of generating text for them.

The researchers essentially made the AI less convenient on purpose.

Rather than saying:

Here's a better paragraph.

the system might push the student:

What evidence supports that claim?

What would someone who disagrees with you say?

Why is this example stronger than the previous one?

The study was tiny—just four students—so it certainly does not establish a universal solution. But the design idea is fascinating. Students using these deliberately constrained systems developed stronger counterarguments and deeper revisions. The researchers described the approach as creating a kind of productive friction around the student's own thinking.

Think about what that means.

The best educational AI may not always be the AI that gives the best answer.

It may be the AI that knows when not to give one.

That turns our usual definition of technological progress upside down.

A consumer AI assistant is rewarded for saving you effort.

An educational AI system may sometimes need to preserve effort.

Schools have always engineered difficulty

We sometimes forget this.

Most classroom activities are artificially difficult.

A teacher could simply tell students what happens at the end of the novel.

We make them read it.

A calculator could multiply the numbers.

Children still learn multiplication.

A translation tool can produce excellent French.

We still ask students to learn another language.

Wikipedia can tell you when the French Revolution occurred.

Students are still expected to know enough history to understand why it mattered.

Education deliberately asks people to do things that technology—or another human being—could do for them.

Why?

Because the product is not the point.

The changed person is the point.

A piano student plays scales even though Spotify can produce far better music.

An athlete lifts weights that accomplish no useful movement of an object from one place to another.

The apparent inefficiency is precisely what develops the capacity.

School has always contained the cognitive equivalent.

And AI makes it necessary to identify those exercises more explicitly.

If an assignment can be completed instantly by AI, educators need to ask:

Was producing the answer the goal?

Or was the struggle required to produce it doing something important?

Sometimes the answer will be no.

Some assignments probably deserve to die.

But others may need to become deliberately harder to outsource.

The strange case of the blank page

Consider brainstorming.

It seems like the perfect use for generative AI.

A student has no idea how to begin.

The chatbot instantly provides ten ideas.

Problem solved.

Except generating those first possibilities may itself be one of the abilities we hoped the student would develop.

A new Common Sense Media survey released this month found that 70% of American teenagers use AI for schoolwork. Among AI users, 38% said access to it causes them to generate fewer of their own ideas, while 39% said they sometimes feel they are missing out on learning when AI completes assignments.

Those numbers interest me because they suggest students aren't oblivious to what is happening.

They can feel themselves outsourcing something.

At the same time, two-thirds said AI helps them understand their schoolwork.

Both things can be true.

The same machine can deepen understanding at 3:00 p.m. and eliminate thinking at 8:00 p.m.

The difference may come down to when it enters the process.

This suggests a surprisingly simple principle.

Think first. AI second.

Before asking AI to generate ideas, generate three yourself.

Before asking AI to solve the problem, attempt it.

Before asking AI to critique an argument, decide what you believe.

Before asking AI to summarize a reading, read enough of it to know whether the summary is any good.

Before asking AI what a character means, form an interpretation worth challenging.

This isn't anti-AI.

It is sequencing.

And sequence may become one of the most important design questions in education.

We need more than "AI literacy"

Schools increasingly say students need AI literacy.

I agree.

But I'm not convinced we have settled what that means.

Often, AI literacy sounds like another technical competency:

How to write a good prompt.

How large language models work.

How to verify sources.

How to detect bias.

How to protect personal information.

All worthwhile.

But there is a deeper form of AI literacy that is less about understanding the machine and more about understanding yourself while using it.

Did I know this before AI told me?

Could I explain this without looking?

Am I asking for help because I am genuinely stuck—or because thinking for another five minutes feels unpleasant?

Did the tool improve my idea or replace it?

Am I persuaded because the argument is good or because it is fluent?

Would I recognize if this answer were wrong?

That is metacognition.

And AI may make it dramatically more important.

A 2026 experimental study involving 226 participants found that interventions explicitly designed to encourage critical thinking during human-AI problem solving reduced people's tendency to directly adopt AI-generated material and produced more original and creative solutions.

Again, the important variable wasn't simply access to AI.

It was the structure surrounding its use.

Perhaps the most important AI skill will therefore not be prompting.

It will be calibrating dependence.

Knowing how much of your thinking to keep.

The child who struggles may need AI most—and be most vulnerable to it

There is another complication.

The children most tempted to outsource work may also be the children who could benefit most from individualized assistance.

Common Sense Media's 2026 census found that children who reported difficulty concentrating, writing essays, learning mathematics or persisting through challenging tasks tended to use AI for schoolwork more frequently. For example, 56% of children who struggled to stay focused said they used AI for schoolwork at least weekly, compared with 45% of those who didn't report that difficulty.

This is where blanket rules become especially crude.

For one student, AI may be a shortcut around learning.

For another, it may be the first tutor who can explain a concept patiently enough for learning to happen.

The same feature can be scaffold or crutch.

And you often don't know which until you understand the learner.

That is a deeply inconvenient fact for education systems designed around universal policies.

It is also one reason teachers become more important, not less.

Someone still needs to know the child well enough to ask:

Is this tool expanding your capacity—or compensating for it so completely that the capacity never develops?

An algorithm cannot easily make that judgment from the final assignment.

A good teacher sometimes can.

The future classroom may have different modes of thinking

Instead of dividing schools into "AI allowed" and "AI prohibited," I suspect we will eventually need more sophisticated categories.

There may be moments of closed cognition:

No AI. No search. No notes.

Think.

Remember.

Calculate.

Write.

Struggle.

We already do this with exams, but the purpose would not simply be surveillance. It would be deliberate cognitive exercise.

Then there may be AI-assisted cognition:

Use the machine, but document what it contributed.

Challenge its answers.

Compare alternatives.

Verify sources.

Explain what you changed.

And finally, perhaps, AI-native work:

Here the student's job is explicitly to accomplish something difficult with artificial intelligence.

Analyze a huge dataset.

Simulate competing scenarios.

Build something.

Interrogate several perspectives.

Produce work that would have been impossible—or prohibitively time-consuming—without the technology.

Children may need experience in all three modes.

Because the adult world they enter will require all three.

Sometimes we will need to think alone.

Sometimes we will need to think alongside machines.

And sometimes we will need to recognize that the machine is thinking so much for us that we have stopped noticing what we no longer know how to do.

This changes the teacher's job too

There is an irony in all the speculation that AI will make teachers less necessary.

If information becomes abundant, the valuable part of teaching shifts.

A teacher no longer needs to be the fastest source of an explanation.

AI will win that contest.

But education was never merely information delivery.

A teacher notices that the student who normally argues passionately has stopped contributing.

A teacher knows that a beautifully written paragraph doesn't sound like the fourteen-year-old sitting in front of her.

A teacher can say:

"I don't want the better answer yet. Tell me what you think."

A teacher can create a room where twenty students disagree with one another and still have to continue the conversation.

A teacher can decide when a child needs help and when the most useful thing is to let the child remain stuck for another sixty seconds.

That last skill may become increasingly valuable.

AI is extraordinarily good at answering.

Good teaching sometimes requires resisting the urge to answer.

Maybe schools should stop trying to detect AI

There will still be legitimate reasons to investigate academic dishonesty.

But an education system organized around catching AI use is probably fighting the technology at the wrong layer.

Students will become better at hiding it.

AI-generated writing will become harder to distinguish.

Detection tools will remain imperfect.

The surveillance burden will grow.

There is another possibility:

Design work whose educational value survives the presence of AI.

Ask students to defend choices.

Require intermediate thinking.

Compare drafts.

Make reflection part of assessment.

Give students AI-generated arguments and ask them to find the flaws.

Let AI produce three solutions and ask which one fails.

Have students improve a bad AI answer.

Ask them to explain something orally after submitting it.

Reward intellectual changes of mind.

Assess questions as well as answers.

Make students show not simply what they concluded, but where their judgment entered the process.

At that point, whether AI was present becomes less interesting.

The student still has to be present.

The real question

AI will probably eliminate some educational tasks.

Good.

Education has accumulated plenty of rituals simply because they were easy to administer.

But we should be careful about confusing an obsolete task with an obsolete ability.

Maybe five-paragraph essays deserve to disappear.

That does not mean forming an argument does.

Maybe memorizing huge quantities of information becomes less important.

That does not mean knowledge does.

Maybe drafting routine prose will increasingly be automated.

That does not mean learning to express a thought clearly becomes irrelevant.

In fact, the opposite may happen.

When competent answers become cheap, judgment becomes expensive.

When polished writing becomes abundant, having something worth saying matters more.

When every student can generate twenty ideas, deciding which idea deserves to exist becomes a skill.

And when information becomes available instantly, knowing enough to recognize nonsense becomes more valuable, not less.

That is why I think we are asking the wrong question about AI in schools.

The challenge is not to decide whether children should grow up using artificial intelligence.

They will.

The challenge is deciding what capacities we want them to develop strongly enough that technology remains a tool they can choose to use rather than an ability they quietly become unable to function without.

Perhaps every assignment in the AI era should ultimately face one test:

What is the student supposed to become capable of doing because we asked them to do this?

If we cannot answer that, AI may not be the problem.

The assignment may be.

And if we can answer it, then we know what part of the work we should never allow the machine to take away.

Artificial IntelligenceeducationNon featured

Nam Nguyen