← All insights

A · Manage · 2026-09-07

As AI Grows Stronger, Why Are Schools Restricting Its Use Instead?

Illustration · not a service record: Workshop in session: participants working with AI toolsIllustration · not a service record

[ As AI Grows Ever Stronger, Why Are New York and Los Angeles Restricting Students’ Use of AI Instead? ]

Last week, two very different pictures appeared in the AI world in the United States.

On one side, OpenAI released GPT-6 Astra and Anthropic launched Claude Fable 5.1. AI took another big step forward: it can complete more complex tasks, and it is more and more like a helper that can get things done on its own.

On the other side, large public school districts on America’s East and West Coasts hit the brakes at the same time.

New York has barred students from early childhood education through eighth grade from using generative AI directly, keeping only limited, supervised use in high school; the Los Angeles Unified School District has suspended the use of generative AI by students in all grades on district-issued devices. This is not a ban on all AI for educational purposes, but it is certainly a sharp tightening of “letting students use AI directly.”

AI has become stronger, yet schools have become more afraid to use it. Why?

I think this echoes exactly what I argued in my previous essay:

A powerful AI is not the same as an AI that makes people powerful.

The former does things for you; the latter helps you build your own abilities.

▍What worries us is not that children get help, but that children lose the chance to practice

Imagine an adult standing beside a child, one who knows everything and is always on call.

Before the child has even thought, the adult gives the answer; when the child gets slightly stuck, the adult finishes the job. The essays read better and the math homework is all correct, but the child’s own chances to practice thinking, judging and revising may become fewer and fewer.

Does it follow that the more capable this adult is, the more capable the child becomes?

Not necessarily. Sometimes the adult’s ability ends up covering over the abilities the child has yet to grow.

But that does not lead us to think that the best education is to send away everyone around the child who could offer help.

What we really need is help that knows when to demonstrate, when to give a hint, when to wait, and when to let go.

The same is true of AI.

Of course, AI cannot replace real teachers and human relationships. But if we are going to let it take part in education, we should not only ask, “How much can it do for students?” We should also ask:

“After it has helped, what abilities has the student gained?”

This is not to argue that the harder learning is, the better. What we want to reduce is unnecessary burden; what we want to keep is the thinking that lets ability grow.

That is why I believe education AI—more precisely, Pedagogical AI designed for teaching and learning—cannot simply be a general-purpose chatbot moved into the classroom.

Its priority should be “making people stronger,” not just “getting things done.”

▍Turning learning science into something AI does from the very start

Last week I was in Frankfurt, in discussions with experts from neuroscience, psychology, education and AI.

As one of only two edtech practitioners in the room, there was one thing I cared about a great deal: can this research, in the end, become product design that teachers can actually use and children truly benefit from?

We discussed more than thirty principles supported by varying degrees of evidence from learning science. I then used three frontier AI models to help compare and organize them, narrowing them down to the nine common directions in the attached figure (see the comments).

Still, the fact that three AIs all arrived at them does not mean science has given its stamp of approval.

The first item in the figure directly cites a learning experiment with generative AI; the other items mainly list existing learning science research. They offer directions worth trying, but once they are put into new tools, different ages and different classrooms, they still need to be tested again.

For example, one high school math study found that students who used a general-purpose GPT tool performed better during practice; but once the AI was taken away, their performance on a test taken on their own was about 17% lower than that of students who had not used it. In the same study, a version with learning-protective design added greatly eased this negative effect.

This does not prove that all AI harms learning, nor does it prove that adding a few lines of teaching prompts will always make students stronger.

But it reminds us that “using AI” is not one single kind of learning experience. How a tool is designed and how students use it can lead to very different results.

For Junyi Academy, these nine principles are not a certification sheet to be used for publicity, but a working list for product design and research.

We are adjusting our products in this direction: so that AI does not just answer faster, but can also ask follow-up questions at the right moment, give hints that are just enough, arrange practice, and gradually step back as students become capable.

And what we must keep asking is not only “Do users like it?” but also “Did they really learn it?”

▍Taiwan does not have to choose between a “total ban” and a “full sprint”

So should Taiwan’s classrooms use AI or not? Do we need to follow with a ban of our own?

My view is this: encourage exploration without forcing anyone on board; keep choices open, and also set a bottom line.

I do not think pausing necessarily means falling behind. When a tool’s safety, age-appropriateness and educational benefit are still unclear, pausing can be a responsible choice.

But if all that remains of a policy is the single word “ban,” without making clear under what conditions trials are allowed, what evidence needs to be accumulated and when the policy will be reviewed, it may also shut valuable exploration out the door.

Likewise, we cannot demand that every teacher and every classroom keep up immediately just because AI is powerful.

Indiscriminate full opening and indiscriminate full blocking both fail to answer the real question in education.

The approach I hope for is to provide a variety of tools that have passed basic safety and pedagogical review, and to let teachers who are prepared and motivated try them first, under suitable conditions of age, curriculum and supervision. Teachers who are not yet ready would have time and support, and would also keep the option of not using them.

“Letting the motivated run first” does not mean having children bear unprotected risks, still less pushing those who are not yet ready into the unknown along with the rest.

Moreover, autonomy must not turn into something where only well-resourced schools have a choice. Training, research and tool support should also look after under-resourced schools.

More important still, pilots should hand in not only usage rates but also evidence of learning.

Are students better able to explain their own thinking? Do they still remember after some time has passed? Can they still apply it when given a different problem? When they need to complete something on their own, can they still do it?

South Korea’s AI digital textbooks went from being pushed hard to being reclassified in 2025 as educational materials that may be used by choice, and this should put us on our guard. The policy turn in itself cannot be taken as proof that “AI education does not work”; but my reading is that the pace at which a system is pushed forward should not exceed the pace at which schools get ready and evidence accumulates.

Taiwan can take a more pragmatic path: pilot first and then expand, able to advance and able to pull back, giving products a chance to adjust and evidence a chance to speak.

▍What can teachers do while the tools are not yet mature?

We do not have to wait until the perfect education AI appears before we can start doing the right things.

The rightmost column of the attached figure turns the nine principles into teaching practices that teachers can take into their own hands. Not all of these practices have to go through students operating AI directly; teachers can first use AI to prepare materials and then lead students through the activities.

1. Have AI test students and question students, rather than answer for them.

Use AI to set questions, press for reasons and challenge ideas. If a student keeps asking for hints all the way through, stop and ask them to explain what they understand so far.

2. Help should have a rhythm, and should also be withdrawn step by step.

The teacher decides when to demonstrate, when to give a hint and when to reduce help. The purpose of a handrail is to help students stand steady, not to let them lean on it forever.

3. Don’t just ask “Do you understand?”; also ask students to say it back.

Have students explain in their own words, teach it to someone else, or talk through their thinking as they solve a problem.

4. A little less rereading, a little more active recall.

Arrange low-pressure quizzes so that students first try to find the answer from memory, rather than keep looking at the answer and nodding.

5. Spread practice out, and bring old knowledge back.

Do not squeeze all the practice into the same day; as new lessons keep moving ahead, keep scheduling cumulative review as well.

6. After initial familiarity, mix different types of problems.

Have students practice judging “which method does this problem call for,” and explain in advance that mixed practice may feel harder, and that this does not mean they have gotten worse.

7. Treat AI feedback as a draft for discussion, not a verdict.

Teachers and students check together which suggestions are reasonable and which need revising, and do not accept everything wholesale just because the AI said it.

8. Break long content into short segments and leave pauses.

Let students pause, digest and look again; do not drown learners in a full page of text or content that plays without stopping.

9. Keep cues consistent, and do not let the screen steal attention.

Use clear structure and fixed cues to help understanding, and reduce decoration and distraction that have nothing to do with learning.

None of these is an iron rule that applies regardless of age or level. For novices especially, first understanding a demonstration may matter more than being asked to think for themselves from the very start.

The teacher’s judgment remains key. But we also cannot throw all the responsibility onto teachers: a good product should build these pedagogical considerations into its design, rather than require every teacher to wrestle with a tool’s default behavior every day.

▍What Taiwan has a chance to contribute is more than just another AI product

This is also why I still have high hopes for Taiwan’s education AI.

But this hope does not need to rest on the assumption that other countries will not do it.

The four national AI missions that Singapore put forward this year focus on advanced manufacturing, transport and logistics connectivity, finance and healthcare, and do not list education separately; but Singapore’s Ministry of Education has already been rolling out AI features with pedagogical design and safety rules on SLS, its student learning platform. So not being listed among these four missions does not mean there is no investment in education AI.

The United States also has organizations such as Khan Academy that keep exploring learning-centered AI. As for products from China or any other country, whether they can reach the world should, I believe, depend on whether they can cross the thresholds of language, curriculum, culture and trust, rather than being decided by nationality alone.

Taiwan’s opportunity is not to wait for others to be absent, but to choose to do one important thing in depth.

The space I really see is this: the AI that most easily brings commercial returns is not necessarily the AI most worth investing in for education.

Helping children build foundations, staying with learners who have fallen behind, tracking long-term outcomes, making public which methods do not work—we cannot simply expect the market to do this work for us automatically.

This is exactly the task that public education, research institutions and nonprofit organizations should take on together.

If Taiwan can get teachers, researchers and product teams to work together, turning problems from real classrooms into design and then turning design into credible evidence, we will have a chance to export not just products, but:

Methods for education AI that different countries can all use, test and improve.

This time in Germany, at an international meeting of fifty people, there were four participants from Taiwan, and I also met a participant’s family member who is from Taiwan. This is of course no proof of any global ranking, but it made me feel very concretely that people from Taiwan can play a role in this cross-disciplinary conversation that exceeds our share of the population.

So if we take “becoming the country that contributes the most to education AI in the world” as a goal, and even drop the words “one of” for the time being, it is not because we have already won, but because this is something worth taking on in earnest.

The world, and China and the United States in particular, is racing to see who can make AI stronger.

Taiwan can at the same time pursue another question: who can make the people who use AI truly stronger?

We must not only bring AI into education, but also bring education into AI.

The breakthrough most worth pursuing in education AI is not to make the next generation need to think less, but to make the next generation more able to think.