Photo by Quilia on Unsplash
South Korea did what nearly every education ministry is currently talking about doing. It put AI-driven textbooks into schools nationally, backed by law and serious money, and it did it first.
Then, sixteen months after the policy was set in motion and five months after the software reached classrooms, the National Assembly passed a bill removing their legal status as textbooks altogether.
That is a rare thing: a full national experiment with a documented ending. Worth reading carefully by anyone about to start one.
What Was Actually Built
Not chatbots bolted onto a PDF. The programme covered adaptive learning software that adjusted difficulty per student, tracked progress, and was designated an official textbook, which in Korea is a legal category carrying budget and adoption obligations.
The money matched the ambition. The Korea Herald reported at least 533.3 billion won, around $385 million, allocated in a single year, against an estimated 800 billion won invested across the sector.
What the Bill Changed
On 4 August 2025 the National Assembly amended the Elementary and Secondary Education Act to define a textbook as printed books and e-books, explicitly excluding “learning support software using intelligent information technology”.
The reclassification is technical and its effect is not. Losing textbook status removed the legal and financial basis for adoption. The software still exists; schools may use it if a principal chooses and finds the budget. Adoption at the time sat at roughly 30 percent.
Photo by Kelly Sikkema on Unsplash
A full national experiment with a documented ending is a rare thing, and worth reading carefully by anyone about to start one.
Lesson One: The Failure Was Operational Before It Was Pedagogical
The debate before launch was about screen time and whether adaptive software helps children learn. The problems on the ground were duller. Lessons delayed by login trouble. Software that did not behave the same way twice. Teachers finding factual errors in material that had been described as vetted.
None of that is an argument about pedagogy. It is an argument about whether the thing works at 8:50 on a Tuesday with thirty children waiting, and that question was answered before the pedagogical one got a hearing.
Lesson Two: Teachers Were Handed a Product, Not a Change
A textbook can be adopted by decree because a book behaves identically in every classroom. Software that adapts, logs and occasionally fails cannot, because it changes what the teacher does minute to minute.
Rolling it out at national scale meant every teacher absorbed the failure modes personally, in front of a class, with no time budgeted for it. Enthusiasm is not the variable there; capacity is.
Lesson Three: Simultaneous Is Not the Same as Fast
The apparent efficiency of a national rollout is that everyone starts at once. The hidden cost is that nobody has been through it before, so there is no school that already knows what breaks and no teacher who can tell a colleague what to expect.
| Staged rollout | National rollout |
|---|---|
| Failures surface in a small group | Failures surface everywhere on the same morning |
| Early schools coach later ones | Nobody has been through it |
| Vendor fixes before scale | Vendor triages at scale |
| Political cost stays contained | Political cost is national from day one |
By the time a reversal was on the table, the story was not “this needs work”. It was already a national grievance with parents attached to it.
Photo by Joshua Hoehne on Unsplash
The question was never whether adaptive software helps children learn. It was whether it works at 8:50 on a Tuesday with thirty children waiting.
Lesson Four: A Policy Tied to an Administration Inherits Its Lifespan
The programme was a flagship of the previous government. Governments change, and initiatives carrying a political signature are re-examined when they do, regardless of what the evidence says by then.
Anything requiring five years of steady support to prove itself needs a foundation broader than one administration’s priorities. That is a design constraint on education technology policy, not a Korean peculiarity.
Lesson Five: A Reversal Is Not a Verdict on AI in Education
The tempting reading is that AI in classrooms was tried and failed. What was tried was one specific approach: proprietary adaptive software, mandated nationally, designated as a textbook, deployed simultaneously.
Plenty of other configurations remain untested. Teacher-facing tools that never touch student data, optional adoption, staged pilots, open materials. The evidence here is against the deployment model rather than against the category, and conflating the two would be the most expensive lesson not learned.
What the money bought
Setting aside whether the policy was right, it is worth seeing where a programme of this size actually spends.
| Where it went | What survives the reclassification |
|---|---|
| Software development by domestic publishers | The products exist and can still be licensed, now school by school |
| Device and network provisioning in schools | Retained. Tablets and connectivity outlive any single programme |
| Teacher training on specific platforms | Largely stranded if a school drops the tool |
| Operational experience | The genuinely durable asset, and the one nobody budgeted for |
What Korea Is Left With
A large domestic edtech sector built around a designation that no longer exists, restructuring accordingly. Schools that may keep the software if they can pay for it. And a body of practical experience that no other country currently has.
Korea also happens to be the first country to bring a comprehensive AI statute into force, which I covered in Korea AI Basic Act, and the same pattern shows there: move first, learn publicly, adjust in law.
FAQ: Frequently Asked Questions
Did South Korea ban AI textbooks?
No. It removed their legal status as official textbooks in August 2025, reclassifying them as supplementary educational material. Schools may still use them, funded and decided locally.
Why did the AI textbook programme fail?
Operational problems came first: login failures, unstable software and factual errors in content. Adoption stalled near 30 percent, and the legal basis was withdrawn before pedagogical questions were settled.
How much did it cost?
At least 533.3 billion won, about $385 million, was allocated in one year, with an estimated 800 billion won invested across the sector.
Does this mean AI does not belong in schools?
It means this particular model, mandated nationally and designated as a textbook, did not hold. Other configurations were not tested here.
Reporting on the legislation and figures reflects coverage as of August 2025 and the situation to 21 August 2026. Education policy in this area continues to change; check current sources before citing any of it.