The Phone Kids Get at Dinner Isn't for Learning

At the restaurant table, the phone in front of the kid was never about learning — it was 30 minutes of quiet the parents were buying

There’s a scene you see all the time on a weekend evening at a restaurant. A phone is propped up in front of a kid who looks about five, and only then do the parents pick up their chopsticks. The phone wasn’t something the kid begged for, and it wasn’t handed over because the parents were being careless. It’s a tool the adults reached for so they could eat in peace for about 30 minutes.

On September 2, 2026, New York City announced a one-year suspension of generative AI use for roughly 600,000 public school students. Because it’s a decision from the largest school district in the US, it was quickly picked up by Korean media, and right behind it came the inevitable chorus: Korea needs to prepare similar regulations.

But if you simply transplant that measure, the restriction only covers classroom instruction — the screen at home stays exactly where it was. In Korea, the screen in a child’s hand has rarely functioned as a learning tool. From the start, it’s been a childcare tool.

New York’s Ban Only Applies Inside School

Let me start by pinning down exactly what New York banned, because a lot of the coverage has blurred the details.

The measures break down into four categories. First, generative AI used directly by students is suspended for one school year, 2026–2027, from 2-K¹ through 8th grade. That covers roughly 600,000 students — two-thirds of the entire student body. Second, companion chatbots² are banned across every grade level. Third, a small pilot opens up five vetted tools to some high schoolers — capped at 50,000 students, up to five classrooms per school, with usage time and assignment counts capped separately for each tool. Fourth, teachers can still use AI for operational tasks like lesson prep or translation, but not for grading or counseling.

Looking at the five approved tools reveals exactly what New York was trying to preserve. Quill supports close reading, text analysis, and finding evidence to back up arguments in English class; Edia assists with math instruction. Brisk Teaching can be applied to teacher-designed activities across any subject. Playlab has students pick apart AI outputs themselves, examining bias and logic. And Intel AI-Ready Schools is project-based — students identify a problem in their community and spend a semester building a solution.

Notice the pattern. None of the five hands students an answer. They either support reading, carry activities that teachers themselves designed, or turn AI itself into the object of analysis. New York City’s own selection criteria pointed the same way: safety and data privacy, a teacher-led classroom structure, and whether students remain the ones doing the thinking. The city clearly wanted to weigh these conditions of use, not just learning outcomes.

On top of this comes a screen-time cap. Students in 2nd grade and below face restrictions on one-device-per-student screen use; grades 3 through 5 get 30 minutes a day, and grades 6 through 8 get 45 minutes. New York City had already banned personal phone use during the school day starting in the fall of 2025, so this stacks one more layer on top of that. Exceptions apply for students with disabilities, English-language learners, and coding classes.

The official AI education every high schooler receives amounts to a 45-minute critical-thinking module, twice a year — 90 minutes total, per year. That’s the entire sum of time 600,000 kids spend learning about AI in school.

The measures themselves are fairly sophisticated. But they apply to exactly one place: inside the school building, during class hours, under a teacher’s watch. New York City has said nothing — and can say nothing — about what a kid does with AI at home.

Korea’s AI digital textbooks saw low adoption in practice

Let’s imagine Korea passed a similar law. First, we’d need to look at how much AI is actually being used in schools. The usage rate for AI digital textbooks has already fallen sharply.

Korea announced the introduction of AI digital textbooks in June 2023 and spent ₩1.4093 trillion (~$1.02 billion) over three years. ₩96.3 billion (~$70 million) went to building wireless networks in classrooms, ₩117 billion (~$85 million) to teacher training alone, and roughly ₩660,000 (~$480) was spent per student. In March 2025, the textbooks were first introduced for English, math, and informatics in elementary grades 3–4, middle school grade 1, and high school grade 1.

You know how it turned out. As the Ministry of Education retreated to a voluntary-adoption policy, the adoption rate stalled at around 33%, and the actual student usage rate disclosed during a parliamentary audit was just 8.1%. The number of schools using the textbooks fell 58.8%, from 4,095 in the first semester of 2025 to 1,686 in the second. With the revision of the Elementary and Secondary Education Act on August 4, 2025, the legal status of these materials was downgraded from “textbook” to “educational material,” and as of 2026 they remain optional “AI educational materials” that schools can choose to adopt or not.

The original plan was far more ambitious. The rollout was scheduled to expand step by step—to elementary grades 5–6 and middle school grade 2 in 2026, middle school grade 3 in 2027, and common high school subjects by 2028. The plan called for adoption across nearly every subject except music, art, physical education, and ethics. Now, only that timeline remains as a plan on paper. In March 2026, the target grades were technically expanded, but since the status had already been downgraded to “educational material,” the expansion carries almost no practical meaning in the field. The certification process has also been suspended, and publishers that had already spent on development are reportedly incurring losses in the tens of billions of won per company.

New York decided to suspend generative AI for students for one year, while in Korea, the number of schools using AI digital textbooks has dropped sharply. The targets and methods differ, but in both cases, the momentum toward expanding AI use in schools hit a wall. And in both countries, the decisive factor wasn’t research on learning outcomes—it was budgets and contracts, in other words, procurement. Korea brought the textbooks in through procurement and pulled them back through procurement; New York halted things through procurement and only afterward said it would study the issue.

So where, then, was children’s screen exposure actually growing? According to the 2025 Smartphone Overdependence Survey released by the Ministry of Science and ICT in March 2026, the proportion of at-risk users³ among the general population fell for the fifth consecutive year, to 22.7%. But the trend diverges by age group. Among youths aged 10 to 19, the figure is 43%; among children aged 3 to 9, it’s 26%—roughly one in four young children.

The at-risk rate for children aged 3 to 9 rose from 17.9% in 2016 to 26% in 2025. This figure alone doesn’t tell us whether the increase in usage happened at school or at home. But it does make one thing clear: we need to separate AI use in schools from overall smartphone use.

There’s one more thing worth noting. It’s not that Korea has made no decisions at all in this space—it’s that AI specifically has been left untouched.

• Since March 2026, a rule banning smartphone use during class has been in effect, with provisions allowing schools to restrict even carrying phones on campus if necessary.
• The Broadcasting and Media Communications Commission included, in its presidential policy briefing, a proposal to restrict social media sign-ups for those under 14 and to regulate recommendation algorithms for those aged 14 to 19.
• In the June 2026 local elections, mandatory parental consent for social media sign-ups by those under 16 emerged as a central campaign pledge, with superintendent candidates voicing similar positions.
• Starting in the second semester of 2026, the Ministry of Education will run “smartphone-free schools” at pilot institutions.

Put it all together, and a sequence emerges. Korea is a country that regulates smartphones and social media first, and AI later. In that gap, conversational AI—the kind children actually use—remains outside the scope of the regulatory conversation.

That Screen Isn’t Education or Entertainment—It’s Childcare

Let’s go back to that phone at the restaurant.

Regulatory debates almost always start from an addiction model: the child craves stimulation, the platform’s design amplifies that craving, and the child can’t stop on their own. The EU’s approach targeting infinite scroll and autoplay stands exactly on this model. And in plenty of domains, that model fits well.

But the phone placed in front of a five-year-old doesn’t fit this picture. The decision-maker in that moment isn’t the child—it’s the adult. The adult judges the situation and hands over the phone, then takes it back once their own need has passed. The addiction model assumes “the child uses it because they want to.” Here, “the adult gives it because they need to” comes first.

Why do they need to? The schedule tells the story. First graders in elementary school typically get out around 1 p.m., while parents’ workday ends much later. Neulbom School—Korea’s extended after-school childcare program—was created precisely to fill this gap. Before it launched, the after-school program utilization rate stood at 50.3%, and the care-classroom utilization rate at 11.5%. The very fact that the state is pouring in hundreds of billions of won and tens of thousands of staff to close this gap shows just how large it is.

The 30 minutes at the restaurant is the same kind of gap, just smaller in scale. The character is identical: a screen filling in for a shortage of adult hands. Seen this way, the screen isn’t so much content the child watches as childcare the parent has outsourced.

Why does it have to be a screen? Because there’s no alternative. A coloring book is finished in 5 minutes, toys make noise, and calling in a grandparent or a babysitter takes money and advance planning. A screen works instantly, costs nothing extra, keeps the child from wandering off, and—above all—doesn’t inconvenience anyone nearby. The more a society tends to judge parents when a child makes a scene in public, the more rational this choice becomes. What the parent gains from the screen isn’t really the child’s enjoyment—it’s freedom from other people’s disapproving looks.

That’s why treating this behavior as an education problem doesn’t work well. Campaigns warning parents about the harms of screens have already been run plenty of times. The fact that the at-risk-for-overdependence rate among young children has kept climbing for 10 years straight isn’t because parents don’t know. It’s because, even knowing, they have no other option in that moment. This isn’t a shortage of information—it’s a shortage of resources.

Why does this distinction matter? Because it completely changes the nature of a ban. Banning an addiction removes one thing a child can no longer do. But banning a substitute for childcare adds one thing an adult now has to do. The former shifts the cost onto the child; the latter shifts it onto the parent.

Why Conversational AI Can Keep a Child’s Attention

Until now, YouTube has mostly filled that role. But video has a limit: it only goes one way. There comes a moment when the child gets bored, or can’t find what they want and calls for a parent, or finishes watching and looks up at a parent again. The quiet time video buys never lasts long.

Conversational AI breaks through that limit. When a child speaks, it answers. Ask the same question ten times and its tone never shifts. It doesn’t get annoyed, doesn’t rush, doesn’t drift off to do something else. It keeps playing along with whatever rules the child invents. A chatbot can sustain, over and over, the kind of responsiveness a tired adult simply can’t keep up.

From the child’s side, too, this is a different kind of thing. A video is something a child watches; conversational AI is something that responds to the child. It can be given a name, it can pick up yesterday’s conversation where it left off, and it follows whatever rules the child sets. It looks similar to a young child talking to a doll, but there’s one decisive difference: with a doll, the child has to invent the doll’s lines themselves, while with AI, the lines come out without the child having to make them up. The part where the child has to conjure the dialogue on their own simply disappears.

For parents, this can actually feel reassuring. Compared to not knowing what YouTube’s algorithm will serve up next, a partner that only answers when the child speaks to it feels more controllable. There’s no provocative content, and few ads. So this shift happens with less guilt attached. Screen performance goes up in a way that doesn’t collide with the norm of “cut down screen time.”

To see why this combination is dangerous, it helps to revisit New York’s decision. New York announced two measures on the same day: an age-based one (a one-year delay for grades 2-K through 8) and an age-agnostic one (a ban on companion chatbots across every grade). The press led with the former in their headlines, but the latter rests on firmer ground. It reflects a judgment that the core problem isn’t the child’s age at all — it’s a product designed to form a relationship with the user, like a friend or conversation partner.

In a previous issue, What Do Teens Actually Talk to AI About?, we covered a US survey finding that 64% of teens were already using chatbots — 16% for everyday conversation, 12% for emotional support. Yet only 51% of parents knew their child was using a chatbot at all. That’s a 13-percentage-point gap from actual usage. Half of adults, in other words, have no idea something is already happening right outside of school.

Can simply blocking access stop this trend? Australia’s results, which we examined in We Tried Banning Teens from Social Media, are instructive here. The Australian government announced it had deleted roughly 4.7 million accounts belonging to users under 16, but a parent survey conducted around the same time found that 69.1% of respondents said their child still had an Instagram account. In other words, the number of accounts deleted and the number of children still holding and logging into accounts are two different things.

And I want to bring back the most uncomfortable finding from that issue. Of 40 randomized controlled trials testing whether quitting social media improves mental health, not a single one had an average participant age under 18. The intervention had never been validated in the exact age group the policy targets. For AI, the evidence base is even thinner than that.

How Far Does This Argument Actually Hold

Let me check for myself how far my argument actually holds up. The “care substitute” explanation doesn’t apply to every age group.

It fits well for preschoolers through the early years of elementary school. The adult is the one handing over the screen, and the adult is the one taking it back. But once kids move past the upper elementary grades, the picture flips. The child has their own device, creates their own account, and uses it during hours their parents don’t know about. The fact that the at-risk-for-overdependence rate among adolescents climbs to 43% isn’t because parents handed them a screen. In this bracket, the addiction model fits far better, and design regulation is most effective when it targets precisely this group.

So, to put it precisely: within the 2-K-through-8th-grade range that New York regulated, two fundamentally different populations are lumped together. In the front half, adults make the decisions; in the back half, children do. A single ban applied to a single age bracket can’t handle that difference.

Let me also put a counterargument on the table. My premise that school use is free and supervised doesn’t always hold. There have been ongoing concerns about how student learning records collected by AI digital textbooks are handled. The learning data of 4.83 million students has been accumulating in both a national database and private edtech companies’ systems, and a separate child-data-protection framework for this hasn’t been built out yet. While the U.S. overhauls its children’s online privacy rules and Europe classifies AI in education as high-risk, Korea’s corresponding safeguards remain thin. So school use being “school use” doesn’t automatically mean it’s safe use — it’s just a different kind of supervision.

Even so, my conclusion stands. School use has, at minimum, a designated party responsible for oversight, and someone to hold accountable if something goes wrong. Use at home and in the private-education market has neither.

If Schools Ban It, Who Absorbs the Burden?

Let’s say a school AI ban actually gets implemented in Korea. Three things happen, in order.

First, there’s almost nothing to regulate. As we saw earlier, student usage of AI digital textbooks was 8.1%. That figure doesn’t capture all AI use in schools, granted, but the point stands: the policy gets announced, yet the actual time removed from a kid’s day is small.

Second, usage at home stays exactly where it was. With 26% of young children in the at-risk-for-overdependence group, evenings and weekends matter just as much as school hours. A regulation confined to school grounds can’t touch the hours logged at home.

Third is the problem unique to Korea. When public education stops using AI, private tutoring fills the gap.

Look at the 2025 survey on private education spending for elementary through high school students: total spending was ₩27.5 trillion (~$19.9 billion), down for the first time in five years. Most of that drop traces to the student population falling 2.3% to 5.02 million. But among students who actually participate in private tutoring, average monthly spending per student rose 2.0% to ₩604,000 (~$437) — crossing the ₩600,000 line for the first time. The share of students spending over ₩1 million a month climbed to 11.6%. Broken down by income, households earning ₩8 million or more a month spent ₩662,000 (~$479), while households under ₩3 million spent ₩192,000 (~$139) — a 3.4x gap. Participation rate fell to 75.7%, even as spending among participants rose. In other words, spending is splitting toward the two extremes.

AI learning tools are entering this market. And cram schools and workbook subscriptions aren’t schools. They sit entirely outside the range of any regulation aimed at schools.

So here’s the outcome you’d expect. After a ban, kids’ AI usage doesn’t drop much. What changes is the character of that usage. Free, supervised, equally-distributed-to-every-child use disappears. What’s left is paid, loosely supervised use, sorted by ability to pay.

On the childcare side, the effect is even more direct. A parent who used to fill 30 minutes at a restaurant with a screen now either spends that 30 minutes actively tending to the child or pays for childcare help. Families who can afford it hire someone; families who can’t just keep handing over the screen, one way or another. What the regulation produces isn’t less usage — it’s stratified usage.

One thing worth flagging here: stratification isn’t only about gaps in how much is used. It’s also a gap in supervision. A household spending ₩660,000 a month on private tutoring and one spending ₩190,000 have access to different tiers of service. They may also differ in how much capacity they have to find someone to sit beside the child while they use AI. These spending statistics don’t directly show how AI is actually being used, but the disparity that would emerge from leaving supervision entirely to individual households is something we need to reckon with. Australia’s social-media ban for minors ran into the same shape of problem in the end. Only the children of households that followed the rules diligently saw reduced access; children of households that didn’t kept right on using it.

This is where the one real advantage of schools comes in. The advantage of school isn’t that it delivers the best education — it’s that it gives every child the same conditions. AI digital textbooks failed for a number of reasons, but at minimum, they were something that could be given identically to all 4.83 million students. Close that channel, and what’s left standing is each family’s ability to pay and its level of information.

From a policy-design standpoint, this is a familiar failure pattern. A ban introduced without a substitute in place doesn’t reduce usage — it just relocates the burden. And that cost always lands first on whoever has the least room to absorb it. If, in designing a regulation, you only ask “who will this stop from doing what,” you’ve only seen half the picture. You also have to ask: “who will this force to do more, and of what?”

Oswarld’s Lens

Where Korea differs sharply from other countries is on the question of age verification.

The wall Australia hit with its under-16 SNS ban was age verification. There was no good way to stop a kid from lying about their age, so even after deleting 4.7 million accounts, usage barely budged. This same limitation is part of why Europe pivoted from age-based blocking to design regulation instead. The turning point was the EU calling TikTok’s infinite scroll illegal.

In Korea, that wall is low. Mobile phone identity verification⁴ is effectively the default for signing up to almost every service. Real names and dates of birth are confirmed through carrier authentication, and that infrastructure has been running for nearly 20 years. In other words, Korea is one of the few countries that can actually enforce this kind of law.

Enforcement capacity is usually seen as a good thing. On this issue, I see it differently. In a country without enforcement power, getting the standard wrong just ends up as a declaration with no teeth. But in a country that can actually enforce, a wrong standard gets applied for real — and reversing it takes time. Korea has already been through this once, with AI digital textbooks. After spending ₩1.4093 trillion (~$1.01 billion) and hitting a utilization rate of just 8.1%, the law was amended to downgrade their status. The rollout was decided through procurement, and the withdrawal was decided through procurement too. Both times, the supporting evidence arrived only after the decision had already been made.

So as I see it, what Korea needs to decide first isn’t whether to ban — it’s whether to regulate based on age, or based on product design.

Age-based regulation is easy to enforce but weak in effect. It’s never been validated against the actual kids it targets, it doesn’t apply outside school, and there’s already a wide-open lane for continued use in private after-school academies (hagwon). Design-based regulation works differently. What gets regulated is the product itself — features engineered to form relationship-like bonds, mechanisms that keep a child hooked when they try to quit, character settings designed to induce emotional dependency. These can be regulated regardless of age, and the workarounds are much narrower. New York’s companion chatbot ban, applied uniformly across all grade levels with no age distinction, is exactly this approach — though the press coverage focused mostly on the one-year grace period for the age-verification requirement.

There’s one more piece. If you’re going to ban a substitute for caregiving, you need to offer a substitute for the substitute. A clause that takes the screen away and a budget line for who fills that time need to live in the same document. A ban without that ends up being filled by parents’ time instead — and that burden falls hardest on households with the least slack to spare.

This isn’t idealism. Korea has already calculated, once, what it costs to fill a caregiving gap. Neulbom School, a state-run after-school care program, was precisely an attempt to fill the after-school care gap for young elementary students using the national budget — it meant assigning dedicated administrative staff to every school and securing more than 10,000 instructors. Filling with people what a screen used to do for free required exactly that much budget and manpower. So when I look at whatever regulation comes next, there’s one question I check first: is there a clause about substitute resources sitting right next to the ban clause? If not, then what that law actually does isn’t reduce a child’s screen time — it just shifts that time onto the parents to fill themselves.

One last thing I want to flag: the pace of this discussion is running backwards. Smartphone and SNS regulation is already at the stage of bills, campaign pledges, and enforcement decrees — yet conversational AI, which would actually be the best-fit substitute for caregiving for a child, hasn’t even entered the conversation yet. In areas where regulation lags, products get there first. And a product that has already staked its claim by the time regulation arrives is far harder to regulate later — because by then it will likely already be the tool filling evening hours in millions of households.

Closing

Let me leave you with three ways to read this issue, depending on where you sit.

If you’re a parent, start with this: the school’s decision doesn’t change how much AI your kid uses. What changes is who’s watching when they use it. If school doesn’t use it, that use just moves home. The question that matters right now isn’t ban-or-not — it’s the fact that half of parents don’t know what their kids are actually using.

If you’re designing policy, keep age thresholds and design standards as separate conversations. Bundle the two into one sentence, and the weaker argument borrows legitimacy from the stronger one. That’s literally what happened in New York.

If you’re building education products, pay close attention to how New York picked its five tools: safety, data privacy, teacher-led use, and design that keeps the student as the one doing the thinking. Going forward, I think these four criteria — alongside learning outcomes — are what any education tool should be measured against.

And the question that’s left standing at the end is this: what do we plan to replace those 30 minutes at the restaurant table with?

💬 Reader, was there a moment you handed a screen to your child, or a niece or nephew, or a kid nearby? Tell us in the comments what that screen ended up doing in your place.

📨 If you’re raising a kid and facing this every day, or if you work in education policy or edtech, please pass this along. It’s worth reading before the ban debate even starts.

References & Further Reading

Primary sources

• New York City Mayor’s Office, “Mayor Mamdani and Chancellor Samuels Put Students First with Nation’s Broadest Generative AI Moratorium in Schools”, September 2, 2026. ··· This is the original text of the policy. The grade-level distinctions and pilot conditions that got blurred in press coverage are laid out here in full.
• GovTech, “NYC Schools Hits Pause on AI, Draws Clear Line on Student Use”, September 2, 2026. ··· The most concrete breakdown of the pilot’s scale and the criteria used to select tools.
• The Kyunghyang Shinmun, “AI textbooks, rushed into classrooms and now in trouble, downgraded to ‘supplementary material’ status”, August 4, 2025. ··· Includes a National Assembly finding that adoption sat at 33% while actual usage came in under 10%.
• Nongmin Shinmun, “The bottomless scroll: 4 in 10 teens show ‘smartphone overdependence’”, March 26, 2026. ··· Breaks down the 2025 smartphone overdependence survey by age group.
• Korea Education Newspaper (Hangyo), “2025 Survey Results on Private Education Spending for Elementary, Middle, and High School Students”, March 2026. ··· The numbers reveal a polarizing structure: total spending fell even as per-participant spending rose.

Background

• National Information Society Agency of Korea, “Smartphone Overdependence Survey”, annual reports. ··· Good primary-source material if you want to check the age-cohort time series directly.
• Newsis, “‘No Instagram under 16?’ — the teen social media ‘brake’ shaking up the local elections”, May 29, 2026. ··· Shows where Korea’s regulatory debate currently sits between Australia-style age gating and EU-style design regulation.

Related past issues

• We Tried Banning Teens from Social Media. Did It Work? ··· Covers two key figures from Australia’s ban and the unresolved problem of never having tested it on under-18s.
• What Do Teenagers Actually Talk About with AI? ··· Maps the gap between adult fears and how teens actually use these tools.
• The Real Reason the EU Called TikTok’s Infinite Scroll ‘Illegal’ ··· Traces where the approach of regulating design rather than age came from.

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: A free public education program New York City offers to 3-year-old children. It corresponds to Korea’s pre-kindergarten program for 3-year-olds. Domestic Korean reporting often translated this as “pre-K,” but the original term is 2-K. ↩

2. Companion chatbot: A chatbot designed to function as a conversation partner, friend, or source of emotional support rather than to provide information or complete tasks. Its core features are a character persona and the maintenance of an ongoing relationship. ↩

3. Smartphone overdependence risk group: Under the Ministry of Science and ICT’s scale, a respondent showing all three of increased salience, failure of use control, and problematic outcomes is classified as high-risk; showing some of these lands them in the potential-risk group. The “risk group” figure combines both. ↩

4. Phone-based identity verification: A method of confirming an online service subscriber’s identity using the real name and date-of-birth information held by telecom carriers. In Korea, it functions as the de facto default authentication method for many services. ↩