NYC Banned AI in Schools. Korea's Screens Are on the Table

On the Restaurant Table, What Parents Bought Was 30 Minutes

There is a familiar scene in restaurants on weekend evenings. A smartphone stands propped up in front of a 5-year-old, and only then do the parents pick up their chopsticks. The phone was not handed over because the child begged, nor out of parental neglect. It is the tool an adult used to buy 30 minutes of quiet.

On September 2, 2026, New York City announced a 1-year moratorium on generative AI for roughly 600,000 public school students. Coming from the largest school district in the United States, the decision quickly made headlines in South Korea, sparking immediate calls that Korea should prepare similar regulations.

Yet if we simply copy that policy, the restriction will stop at the classroom door while the child’s screen remains untouched on the dining table. In Korea, the screen in a child’s hands was rarely an educational tool to begin with. It was an instrument of childcare from the start.

The Line New York Drew Was the ‘Classroom’

Let’s start by looking precisely at what New York actually banned. The media coverage glossed over quite a few details.

The measures fall into 4 main categories. First, direct student use of generative AI is suspended from 2-K¹ through Grade 8 for the 1-year 2026–2027 school year. This affects about 600,000 students, or 2/3 of total enrollment. Second, companion chatbots² are banned across all grades, without exception. Third, only a subset of high school students will be granted pilot access to 5 vetted tools, capped at 50,000 students, up to 5 classrooms per school, and 10 to 20 minutes per tool per week. Fourth, while teachers can use AI for administrative tasks like lesson prep and translation, they are barred from using it for grading and student counseling.

Looking at the 5 permitted tools reveals what New York was trying to preserve. Quill aids close reading, text analysis, and evidence-gathering in English classes, while Edia assists with math instruction. Brisk Teaching can be applied across subjects to teacher-created activities. Playlab lets students dissect bias and reasoning inside AI outputs, and Intel AI-Ready Schools is a semester-long, project-based program where students identify local community problems and build solutions.

Notice the common thread: not one of these 5 tools generates answers on the student’s behalf. They support reading, deliver teacher-designed activities, or turn AI itself into an object of critical analysis. The selection criteria announced by New York City pointed in the exact same direction: safety, data privacy, teacher-led instructional structures, and ensuring the student remains the primary agent of thought. Boosting learning outcomes was nowhere on the criteria list.

Screen-time caps were layered on top. 1-to-1 device use is banned entirely for Grade 2 and below, capped at 30 minutes a day for Grades 3 through 5, and 45 minutes for Grades 6 through 8. New York City had already banned personal smartphones during the school day starting in Fall 2025, effectively adding another layer of restriction. Exceptions were carved out for students with disabilities, English language learners, and coding classes.

One number stands out. The official AI education provided to all grades consists of a 45-minute critical-thinking module twice a year—just 90 minutes annually. That is the total amount of time 600,000 children will spend learning about AI at school.

The policy itself is fairly sophisticated. The criteria for choosing the 5 tools prioritized safety, data privacy, teacher-led instruction, and keeping students at the center of their own thinking. Yet every single one of these lines is drawn within one place: school, during class hours, under a teacher’s watchful eye. As for what a child uses at home, New York City said nothing—and can say nothing.

Drawing That Line in Korea Is Like Padlocking an Empty Room

Let’s imagine passing a similar law in Korea. The first hurdle is that there is hardly anything left to regulate.

In June 2023, Korea announced the rollout of AI digital textbooks, committing ₩1.4093 trillion (~$1.01B) over 3 years. That included ₩96.3 billion to install classroom wireless networks and ₩117 billion solely for teacher training—roughly ₩660,000 (~$470) per educator. In March 2025, the software made its debut in 3rd–4th grade elementary, 1st-year middle school, and 1st-year high school classes for English, mathematics, and informatics.

The outcome is well known. Once the Ministry of Education stepped back to make adoption optional, the adoption rate stalled at around 33%, and the actual student utilization rate disclosed during the parliamentary audit was a mere 8.1%. The number of participating schools fell from 4,095 in the 1st semester of 2025 to 1,686 in the 2nd semester—a 58.8% plunge. On August 4, 2025, an amendment to the Elementary and Secondary Education Act downgraded their legal status from official “textbooks” to “instructional materials,” leaving them in 2026 as discretionary “AI learning aids” that individual schools can take or leave.

The original roadmap was vastly more ambitious. It called for expanding into 5th–6th grade elementary and 2nd-year middle school in 2026, 3rd-year middle school in 2027, and common high school subjects by 2028—blanketing nearly every subject except music, art, physical education, and ethics. Today, that timetable exists only on paper. Although the target grades were nominally broadened in March 2026, the downgrade in legal status rendered the expansion virtually meaningless on the ground. Official screening procedures ground to a halt, and publishers who had already committed massive development budgets are now reporting losses in the tens of billions of won per company.

New York spent 1 year purging AI from its classrooms, but Korean schools had already finished the job themselves. The difference lay entirely in the mechanism: New York pulled the plug through policy, while Korean classrooms pulled it by simply refusing to use the tools. In both cases, the decisions turned not on pedagogical research into learning outcomes, but on budgets, contracts, and procurement. Korea ushered AI in through procurement and scrapped it through procurement; New York froze it through procurement and only then promised to study its effects.

Where, then, was children’s screen time actually accumulating? According to the 2025 Smartphone Overdependence Survey published by the Ministry of Science and ICT in March 2026, the overall proportion of users in the overdependence risk group³ declined for the 5th consecutive year to 22.7%. Yet the trend lines diverge sharply by age cohort. The risk group rate stands at 43% for adolescents aged 10–19, and 26% for toddlers and young children aged 3–9—fully 1 in 4 young children.

These numbers did not originate in the classroom. Even as schools sidelined AI textbooks and collected mobile phones at the door, the risk rate among young children climbed relentlessly. Among children aged 3–9, the high-risk cohort grew from 17.9% in the 2016 survey to 26% within 10 years. The trajectory moved almost entirely independent of school policy.

There is one more crucial distinction to make: Korea has not been inactive in this regulatory arena. It simply has not regulated AI.

• Since March 2026, regulations barring smartphone use during class hours have been in effect, granting schools the authority to restrict possession on campus when necessary.
• The Broadcasting, Media, and Communications Commission included restrictions on social media sign-ups for children under 14, along with algorithmic recommendation curbs for users aged 14–19, in its presidential policy briefing.
• In the June 2026 local elections, mandatory parental consent for social media users under 16 emerged as a core campaign pledge, echoed by candidates for metropolitan and provincial education superintendents.
• Starting in the 2nd semester of 2026, the Ministry of Education is rolling out pilot “smartphone-free schools” across designated model institutions.

Lay these pieces side by side, and the sequence becomes clear. Korea is a country poised to regulate smartphones and social media first, and AI later. What unfolds during that intervening lag is the real question this piece seeks to explore.

That Screen Is Neither Education Nor Entertainment—It Is Childcare

Let’s return to that smartphone in the restaurant.

Regulatory debates almost always presuppose an addiction model: a picture where the child craves stimulation, the platform amplifies that desire by design, and the child consequently cannot stop on their own. The European Union’s approach targeting infinite scroll and autoplay rests squarely on this framework. And to be fair, there are domains where that model fits quite well.

Yet the phone placed in front of a 5-year-old does not fit this picture. In that moment, the decision-maker is not the child, but the adult. The adult assesses the situation and hands it over, then reclaims it once their own need has passed. While the addiction model assumes that “the child uses it because they want to,” here the reality is that “the adult hands it over because they need to.”

Why do they need to? Look at the schedule and you have your answer. A 1st-grade elementary student gets out of school around 1:00 PM, while parents finish work much later. Neulbom School (Korea’s expanded public after-school childcare program) was created precisely to bridge this gap; before its rollout, participation in elementary after-school programs stood at 50.3%, and elementary care classes at 11.5%. The very fact that the state is pouring hundreds of billions of won and tens of thousands of personnel into closing this divide illustrates just how massive the void really is.

The 30 minutes spent at a restaurant represent the exact same type of gap. The scale is smaller, but the nature is identical. The screen fills in for the hours when an adult’s hands are simply unavailable. From this perspective, the screen is not content consumed by the child, but childcare labor outsourced by the parent.

Why a screen, of all things? Because there are no viable alternatives. A coloring book lasts 5 minutes, toys make noise, and asking grandparents or calling in childcare help requires money and scheduling. A screen works immediately, incurs no marginal cost, keeps the child firmly in their seat, and above all, prevents any disturbance to those nearby. In a society where a noisy child in a public space draws judgmental glares at the parents, this choice becomes entirely rational. What the parent is buying is not the child’s amusement, but an exemption from social scrutiny.

That is why framing this as an educational problem fails to solve it. We have already had more than enough public awareness campaigns warning parents about the harms of screen time. The proportion of young children at risk of digital overdependence has climbed steadily for 10 years not because parents do not know better, but because in that precise moment, they have no other options. It is not a problem of lacking information; it is a problem of lacking resources.

This distinction matters because it completely changes the nature of a ban. When you ban an addiction, you eliminate one thing a child is allowed to do. But when you ban a childcare substitute, you add one more task that an adult must shoulder. The cost of the former falls on the child; the cost of the latter falls on the parent.

AI Does This Job Far Better Than YouTube

This is the heart of the matter.

Until now, YouTube mostly filled that role. But video has built-in limits: it is strictly one-way. Children eventually grow bored; when they cannot find what they want, they call for their parents, and once the video ends, they turn right back to them. The quiet bought by a screen always came with an expiration date.

Conversational AI shatters that limit. When a child speaks, it responds. It keeps the exact same tone even after hearing the same question 10 times. It never gets irritated, never rushes, and never gets distracted by other tasks. It effortlessly plays along with whatever rules the child makes up. The machine provides, at zero marginal cost, conditions that an exhausted adult simply cannot offer.

The dynamic is fundamentally different from the child’s perspective as well. A video is something to watch, but conversational AI is a partner that reacts. The child can give it a name, pick up yesterday’s conversation where they left off, and set the ground rules. While it may look similar to a toddler talking to a doll, there is one decisive difference: with a doll, the child has to invent every line of dialogue, but AI generates its own lines unprompted. The cognitive labor of imagination vanishes entirely.

For parents, this can perversely feel more reassuring. Compared to not knowing what the YouTube algorithm might serve up next, an entity that simply answers when spoken to seems much more controllable. There are no provocative videos and fewer ads. As a result, this transition happens with far less parental guilt. The screen becomes vastly more capable—all without clashing with the social norm of reducing screen time.

You can see why this combination is so dangerous by revisiting New York’s decision. On the very same day, New York announced both an age-based policy (a 1-year moratorium from 2-K through 8th grade) and an age-agnostic policy (a blanket ban on companion chatbots across all grades). The media ran with the former in their headlines, but the stronger rationale lies with the latter. Their assessment was that the core issue is not an age group, but an entire class of products engineered to simulate human relationships.

In a US survey covered in our previous issue, What Do Teens Actually Talk About with AI?, 64% of teenagers were already using chatbots—16% of them for everyday conversation and 12% for emotional support. Yet only 51% of parents knew their children were using chatbots at all. That is a 13-percentage-point gap between actual usage and parental awareness. It means half of adults have no idea what is already happening outside school walls.

Can blocking access stem this tide? Australia’s experience, which we examined in What Happened When Australia Banned Social Media for Teens, is instructive. The Australian government announced it had deleted roughly 4.7 million accounts belonging to users under 16, but a parent survey around the same time revealed that 69.1% of respondents said their child still had an Instagram account. The number of deleted accounts and the number of active kids were answering entirely different questions.

And let me reiterate the most unsettling finding from that issue: among 40 randomized controlled trials testing whether quitting social media improves mental health, not a single study had a sample with an average age under 18. The intervention had never once been validated in the very age group the policy targeted. When it comes to AI, the empirical foundation is thinner still.

How Far Does This Argument Hold?

Let me test the limits of my own argument. The framing of screens as a childcare substitute does not apply across every age group.

It fits preschoolers and early elementary students remarkably well. In those years, adults are the ones handing over the screen and taking it back. But once children enter upper elementary school, the dynamic flips. Kids get their own devices, create their own accounts, and use them outside their parents’ sight. The fact that the proportion of adolescents at risk of overdependence has climbed to 43% is not because parents handed them screens. For this older cohort, the addiction model is a much better fit, and design regulations are most effective when aimed precisely here.

Put plainly: the span from 2-K to Grade 8 that New York regulates lumps together two fundamentally different groups. In the younger half, adults make the decisions; in the older half, children do. A blanket ban across a single age bracket cannot account for this difference.

Let me also acknowledge a counterargument. My premise that in-school screen use is free and supervised does not always hold. Concerns have repeatedly been raised about how student learning histories collected by AI digital textbooks are handled. Under this setup, the learning records of 4.83 million students accumulate in both national databases and private edtech company servers, without a dedicated child data protection framework yet in place. While the US updates its children’s online privacy regulations and Europe classifies AI in education as high-risk, South Korea’s safeguards remain thin. School use is not inherently safe use; it simply involves a different kind of oversight.

Yet my core conclusion stands. In school, there is at least a designated entity responsible for oversight, and somewhere to assign accountability when things go wrong. At home and in the private education market, even that does not exist.

Who Gets the Bill for the Ban?

Suppose a school AI ban is actually enforced in Korea. 3 things happen in sequence.

First, almost no aggregate usage is regulated. Student-facing AI use in schools was already down at 8.1%. A policy gets announced, but the time actually erased from a child’s day is negligible.

Second, home usage remains entirely untouched. The hours that created the 26% overdependence risk group among toddlers and young children are evenings and weekends, not school hours. Regulating schools cannot reach these windows.

Third, there is a uniquely Korean dynamic: when public education turns it off, private tutoring turns it on.

According to the 2025 survey on primary and secondary private education spending, total expenditure fell to ₩27.5 trillion (~$19.8 billion), dropping for the first time in 5 years. That was largely driven by a 2.3% decline in the student population to 5.02 million. Yet among students who actually participate in private education, monthly average spending per student rose 2.0% to ₩604,000, crossing into the ₩600,000 range for the first time. The share of students spending over ₩1 million per month also expanded to 11.6%. Broken down by income bracket, households earning over ₩8 million per month spent ₩662,000, compared to ₩192,000 for households earning under ₩3 million—a 3.4-fold gap. While the participation rate fell to 75.7%, spending per participant rose, signaling that outlays are polarizing at the extremes.

AI learning tools are flooding into this private education market. And hagwons (private cram schools) and home-study workbook publishers are not schools. They sit comfortably outside the crosshairs of regulations targeted at the classroom.

The outcome is predictable: after a ban, children’s AI screen time barely drops. Instead, the nature of usage changes. What disappears is the usage that was free, supervised, and given equally to every child; what remains is usage that is paid, loosely monitored, and segregated by ability to pay.

In childcare, the impact is even more direct. Parents who used to buy 30 minutes at a restaurant with a screen must now pay for those 30 minutes with their own physical labor or hire someone else. Households that can afford human care hire it; those that cannot keep leaning on screens however they can. What regulation produces is not a reduction in use, but a stratification of use.

One thing deserves special emphasis here: stratification does not mean merely a gap in volume. It is also a gap in supervision. A child from a household spending ₩660,000 a month uses vetted tools in an environment with dedicated instructors; a child from a household spending ₩190,000 uses free apps alone until parents return from work. They use the same technology, but whether an adult is beside them diverges completely. This was precisely the fallout from Australia’s social media ban: access fell only for children from rule-abiding households, while remaining unchanged for everyone else.

Here lies the singular merit of schools: schools are not places that do things exceptionally well, but places that give every child the same baseline. The AI Digital Textbook initiative failed for many reasons, but it was, at the very least, an artifact that could be delivered identically to all 4.83 million students. Shut that channel, and all that remains is each household’s ability to pay and access to information.

From a policy design standpoint, this is a familiar mode of failure. A ban without an alternative is not a ban—it is a cost transfer. And that cost invariably lands first on those with the fewest resources. When designing regulation, looking only at “who is prevented from doing what” captures only half the picture. One must also ask: “who is forced to do what extra work as a result?”

Oswarld’s Lens

Here, I want to point out one decisive way South Korea differs from other countries.

The wall Australia hit with its under-16 social media ban was age verification. When kids lie about their age, there is no real way to stop them; despite purging 4.7 million accounts, usage rates barely budged. This limitation was also what pushed Europe to pivot from age gating to design regulation. The turning point was the EU declaring TikTok’s infinite scroll illegal.

In South Korea, that barrier is low. Mobile phone identity verification⁴ is the de facto default for signing up for virtually any online service. Real names and dates of birth are authenticated through telecom carriers, and that infrastructure has been running for nearly 20 years. In other words, South Korea is one of the few countries that can actually enforce such a law.

Enforcement capability is usually seen as a virtue. On this issue, I view it as the opposite. In countries without enforcement capability, drawing the wrong line ends as mere rhetoric. But in a country with enforcement capability, a misplaced line actually gets drawn. And rolling it back takes time. South Korea has already experienced this once with its AI digital textbooks. After spending ₩1.4093 trillion (~$1.01 billion) only to see usage stall at 8.1%, the government amended the law to downgrade their status. Adoption was decided through public procurement, and retraction was decided through public procurement. In both instances, evidence arrived only after the decision had already been made.

That is why, as I see it, what South Korea needs to decide first right now is not “whether to ban.” It is where to draw the line.

A line drawn at age is easy to enforce, but weak in effect. It has never been validated among the children it targets, its reach does not extend outside schools, and a bypass route through private tutoring (hagwon, private cram schools) is already wide open. By contrast, a line drawn at design works differently. Products engineered to mimic relationships, features designed to hook children when they try to stop, and personas crafted to induce emotional dependency. These can be regulated regardless of age, and their bypass routes are narrow. New York’s quiet, all-grades ban on companion chatbots is precisely this kind—even if public attention was drawn elsewhere.

And there is one more thing. If you are going to ban a childcare substitute, you must provide a substitute alongside it. A clause removing screen time and a budget allocating who will fill those hours must exist within the very same document. A ban without that is an invoice billed directly to adults’ time—and that invoice cannot choose who it gets billed to.

This is not idealism. South Korea has already run those numbers once. Neulbom School—the national after-school care program designed to fill childcare gaps for lower elementary grades with state funding—required assigning dedicated administrative personnel to every school and securing over 10,000 instructors. That was what it cost to replace with human labor what a single screen had been doing for free. So when evaluating upcoming regulatory proposals, there is only one question I plan to look at first: Is there a clause for alternative resources attached right next to the ban clause? If not, what the law actually does is not reduce children’s screen time, but siphon away parents’ time.

Finally, there is one last point I want to address: the speed of the current debate is moving in reverse. Regulations on smartphones and social media have already advanced to bills, campaign pledges, and enforcement decrees, while conversational AI—which could become the most potent childcare substitute for children—has not even entered the conversation yet. Where regulation arrives late, the market arrives first. And products that entrench themselves in the meantime become far harder to regulate later, because by then, they will already be propping up the evening hours of millions of households.

Closing

Here are 3 takeaways on how to interpret this issue from where you stand.

If you are a parent, look first at the fact that a school’s decision does not change how much AI a child uses. What changes is not the volume of usage, but the locus of supervision. When schools step back, the home ultimately fills the void. What matters right now is not whether it is banned, but the fact that half of parents have no idea what tools their children are actually using.

If you shape policy, keep the debate over age thresholds separate from design standards. When you lump the 2 into a single sentence, the weaker argument borrows legitimacy from the stronger one. That is precisely what happened in New York.

If you build educational products, pay close attention to the criteria New York used to select its 5 tools: safety, data privacy, teacher agency, and architectures that keep students as the active drivers of their own thinking. Rather than marketing claims about learning efficacy, these 4 pillars are far more likely to become the procurement rubric of the future.

And the final question that remains is this: What do we plan to replace those 30 minutes at the restaurant with?

💬 Reader, have you ever handed a screen to your own child, a niece or nephew, or a child around you? Tell us in the comments what job that screen was stepping in to do for you.

📨 Please share this piece with parents confronting this dilemma every day, or with colleagues working in education policy and edtech. It is a perspective worth reading before the rush to ban takes over.

References & Further Reading

Primary sources

• NYC Mayor’s Office, “Mayor Mamdani and Chancellor Samuels Put Students First with Nation’s Broadest Generative AI Moratorium in Schools”, September 2, 2026. … The original policy announcement. It contains the exact grade-band divisions and pilot criteria that were blurred in mainstream news coverage.
• GovTech, “NYC Schools Hits Pause on AI, Draws Clear Line on Student Use”, September 2, 2026. … The most detailed breakdown of pilot program sizing and tool evaluation standards.
• Kyunghyang Shinmun, “AI Textbooks Rushed into Classrooms Downgraded to ‘Supplementary Materials’”, August 4, 2025. … Includes National Assembly audit findings highlighting the 33% school adoption rate and under-10% active login rate.
• Nongmin Shinmun, “Trapped in the Infinite Scroll: 4 in 10 Korean Teens Show ‘Smartphone Overdependence’”, March 26, 2026. … A detailed statistical breakdown by age group from the 2025 Smartphone Overdependence Survey.
• The Korea Education Newspaper (Hangyo), “2025 Survey on Private Education Expenditures in Primary and Secondary Schools”, March 2026. … Captures the polarized reality in hard numbers: total spending fell, but spending per participating student rose.

Background

• National Information Society Agency (NIA), “Smartphone Overdependence Survey”, annual reports. … The definitive primary data source for tracking time-series trends across demographic groups in Korea.
• Newsis, “Instagram Ban Under 16? ‘Youth Social Media Breaks’ Shake Local Elections”, May 29, 2026. … Traces how South Korea’s policy debate is navigating between Australia’s blanket age bans and the EU’s design-level safety mandates.

Related past issues

• We Tried Banning Teens from Social Media. Did It Work? … Explores the dual metrics behind Australia’s ban and the fundamental lack of age-appropriate design verification for under-18s.
• What Do Teens Actually Talk About with AI? … Analyzes the sharp divide between adult anxieties and how teenagers genuinely interact with generative tools.
• Why the EU Really Called TikTok’s Infinite Scroll ‘Illegal’ … Unpacks the regulatory philosophy of holding product architecture accountable rather than relying on age restrictions alone.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

1. 2-K: New York City’s free public education program for 2-year-olds (age 3 in traditional Korean age reckoning), equivalent to Korea’s 3-year-old early childhood curriculum. While frequently rendered as pre-K in Korean media, the official text specifies 2-K. ↩

2. Companion chatbot: An AI conversational agent engineered not for informational queries or task completion, but to serve as a conversational partner, friend, or emotional confidant. Persistent character personas and ongoing relationship maintenance are its core design mechanisms. ↩

3. Smartphone overdependence risk group: In the South Korean Ministry of Science and ICT’s assessment framework, individuals exhibiting all three clinical indicators—salience, loss of self-control, and serious adverse consequences—are categorized as high-risk, while those exhibiting a subset are classified as potential-risk. The total risk group combines both tiers. ↩

4. Mobile phone identity verification: A South Korean digital identity verification system that validates an online user’s legal name and date of birth against carrier-registered subscriber records. It functions as the de facto default authentication standard across most domestic web and app services. ↩