What should education teach in the AI era? — the 1.4 trillion won lesson and the conditions for a new Renaissance

While AI tutors raise learning outcomes by 23%, 70% of students feel their own thinking is getting weaker. Korea was the first in the world to pour 1.4 trillion won into AI digital textbooks, only to see the effort collapse within four months. The problem was not the technology. The problem was bringing in the tools before answering the question 'what should we teach?'

---

Background / why this topic now

I ended Part 1, on jobs and the labor market, this way: 'If the education we received until now was education for labor, what kind of education should we give our children?' At the point where AI has begun to replace labor itself, this question is no longer a story about the future.

As of 2026, the global AI education market is worth about $8.35 billion and growing 31–36% a year (Grand View Research, HolonIQ). More than 60% of teachers already use AI in class, and 30% of K-12 students use AI tools daily. China and the UAE have made AI education mandatory at the national level (Stanford HAI 2026).

Yet in the same period, 95% of US college faculty worry about students' overdependence on AI and the decay of critical thinking (AAC&U, January 2026), and 70% of students themselves perceive that their thinking is getting weaker (RAND Corp, March 2026). Korea, the first in the world to roll out AI digital textbooks nationally, effectively withdrew them along with a dismal 8.1% usage rate.

The core of this contradiction is simple: it comes from trying to layer AI on top of the existing way of educating. If you change only the tool without redefining the purpose and structure of education, you get not a utopia but a humanity that does not think.

---

Key data and current state

AI tutor results — the case for utopia

• Khanmigo (Khan Academy): 23% faster acquisition of algebra concepts (SRI International, 2025). Users surged from 40,000 to 700,000 in a year
• Carnegie Learning MATHia: After one school year of use, an average 12-percentile advantage over peers on standardized tests (data from 1 million+ students, 2025)
• Enhanced AI tutors: 127% learning improvement vs. 48% for general chatbots
• Teacher effect: 55% saw improved student outcomes, 44% saw reduced workload (Forbes Advisor)

Decay of thinking — the case for dystopia

• Gerlich study (2025, SBS Swiss Business School): Negative correlation between frequency of AI use and critical thinking. High AI dependence and low thinking scores among 17–25-year-olds
• MIT Media Lab: The AI-using group recorded the lowest brain engagement at the neural, linguistic, and behavioral levels
• RAND Corp (March 2026): 70% of middle and high school students themselves perceive that 'AI is weakening my critical thinking' (up 22 percentage points in 10 months, from 48% in May 2025)
• AAC&U (January 2026): 95% of college faculty worry about students' AI overdependence and decay of critical thinking

Korea's AI digital textbooks — a 1.4 trillion won experiment

• Total spent: About 850 billion won from the government plus about 580 billion won from private publishers = about 1.4 trillion won
• Board of Audit and Inspection findings (December 2025): 8.1% of students used it 10 or more days, 60% never logged in, national adoption rate 32.3%
• Subscription cost estimate: 336.1 billion won in 2025 → 1.0732 trillion won in 2028 (about 2.8 trillion won over four years)
• Outcome: Downgraded to 'educational material' by a legal amendment in August 2025 → effectively withdrawn

? Interpretation: The learning outcomes of AI tutors are being empirically confirmed. The problem is not AI itself but that the tools were brought in without redefining the purpose and structure of education. Korea's 1.4 trillion won experiment delivered that lesson at the highest possible price.

---

In-depth analysis

Don't layer AI onto existing education; redefine the education system

AI is no longer a matter of choice. 82% of students (college) and 58% (high school) already use AI, and only 7% of schools worldwide have AI guidelines (UNESCO). Banning AI in the classroom is like holding back the tide with your hands.

Then the direction is clear. Rather than grafting AI onto the existing way of educating, we must redefine the human system of education and knowledge itself. Professor Chris Dede of the Harvard Graduate School of Education put it precisely: 'If you educate people for what AI does well, you are preparing them to lose to AI. If you educate for what AI cannot do, you get Intelligence Augmentation.'

The OECD, in its 2026 digital education outlook, likewise proposed a three-stage approach: 'without GenAI → with educational GenAI → with general-purpose GenAI', stressing that the key is to build independent thinking and foundational skills first.

Author's view: It is time to move away from training to find the right answer. Mathematical thinking, humanistic reflection, critical questioning: these are areas AI cannot take over. Rather than one-dimensional education that teaches how to use AI, we should prioritize education that teaches students to question and question again and to put their own thinking into writing.

The structure of 29 trillion won in private tutoring — the tools change, the structure stays the same

Korea's private education spending for elementary, middle, and high school students was about 29.2 trillion won in 2024, with a participation rate of 80%. There is optimism that AI tutors will replace private tutoring, but reality is flowing the other way. AI coding academies, AI prompt academies, and AI-assisted admissions consulting are emerging as new areas of private education. Private education spending fell to 27.5 trillion won in 2025, but essay writing, arts and physical education, and career counseling actually increased.

Unless the admissions system changes, only the tools change while the structure stays the same. As long as the CSAT and school grades remain tests of picking the right answer, AI is reduced to a tool that finds the right answer more efficiently.

Author's view: Universities should change from places for careers and employment into places purely for pursuing scholarship and asking questions. But I also know how difficult that transition is, because the entrenched structure is solid. Even so, discussing AI education without fundamental change to the admissions system is like hanging wallpaper in a house without a roof.

 

Academic cheating — change the assessment, not the surveillance

AI-related cheating rose about 370%, from 1.6 cases per 1,000 students (2022) to 7.5 (2026). In the UK alone, 7,000 cases were detected in 2025, and 88% of US students said they had used AI in some form on an assessment.

But there is an interesting piece of data. 94% of AI-generated output goes undetected (University of Reading). And AI detection tools show a 61.2% false positive rate for non-native English speakers (AllAboutAI). For native speakers it is 5.1%, meaning a student whose first language is not English is 12 times more likely to be flagged for cheating even without using AI. The total amount of cheating shows little difference between 2012 (17%) and 2026 (18%); only the method has changed.

Meanwhile, at schools that redesigned their assessments, AI-related cheating fell 40%. That is evidence that changing assessment is more effective than tightening surveillance.

Author's view: If we have neither the ability nor the time to detect AI, we should change the assessment itself. Instead of assignments where students write reports at home, what if we let them use AI fully for gathering materials and basic research, while exams and assessments are written by the student in person, on paper or on a computer, on site? We also need classes built on discussion and questioning. It is a shift to a structure that evaluates not bringing an answer but questioning and interrogating the answer.

 

Whose data are the children's data?

The most overlooked problem left by Korea's AI digital textbooks is data sovereignty. The AIDT collected each student's learning history, error patterns, progress rate, and login times in real time, but the Personal Information Protection Commission pointed out that KERIS's processing policy omitted the items collected and the retention period. The US blocked commercial use of children's data by edtech companies with the April 2026 COPPA revision, and the EU classified educational AI as 'high risk'. Korea alone is growing its edtech market without a dedicated protection framework for children's learning data.

---

Implications for Korea

The 1.4 trillion won lesson — consensus comes before technology

The failure of the AI digital textbook was partly a technology problem, but more fundamentally it was rash to introduce AI while the basic structure of education remained unchanged. The Board of Audit and Inspection's finding of 'full rollout without a pilot, no established technical standards, unverified effectiveness' speaks to technical haste, but the bigger problem was the absence of social consensus.

Author's view: A process of first building social consensus on the direction of education, politically, was needed. Deciding 'which tools to use' before answering 'what kind of education to provide' was the core cause of failure. It seems trial and error and confusion over the direction of education will continue for more than a decade. But if we give up on this process, the educational dystopia we genuinely fear will become real, and humanity risks degenerating into a humanity that does not think.

Educators must be ready too — the system alone is not enough

Change in education is hard to bring about through the system or the efforts of one or two people. Not only children but educators must be ready. Currently 65% of teachers do not use generative AI, and 9% do not know it exists. 76% of leaders believe 'educators have received AI training', but 45% of frontline teachers and 52% of students say they 'have never been trained' (Microsoft, 2026). The American Federation of Teachers (AFT) launched an AI training program for 400,000 teachers because it recognized this gap.

Author's view: Even at enormous social cost, we must create many failures and cases. It is truly hard, but a country that fails to solve this problem will inevitably become dependent on global platform companies, and that dependence will be irreversible. We have a responsibility to give our children opportunities to think, continually.

---

Outlook and variables to watch

• Positive variable — the evolution of AI tutors: Khanmigo users grew 17-fold in a year and are projected to pass 1 million in the 2025–2026 school year. A free tier could expand access to education
• Risk variable — accelerating decay of thinking: In the RAND survey, students' concern about their thinking jumped from 48% to 70% in 10 months. If the trend continues, an irreversible generational cognitive shift could occur within two to three years
• Checkpoint 1: The Korean Ministry of Education's follow-up AI education policy — what the alternative is after the AIDT's withdrawal. Results from the 1,900 pilot schools
• Checkpoint 2: Discussion of admissions reform — how to assess AI literacy or critical thinking in the CSAT and school grades
• Checkpoint 3: Legislation protecting children's learning data — whether Korea establishes a children's data protection framework on the level of US COPPA and EU GDPR

---

Conclusion — the conditions for a new Renaissance

First, we must redefine the purpose of education. Teaching what AI does well only produces people who lose to AI. We must move away from education that rewards the right answer toward education that questions, thinks critically, and expresses one's own thinking in writing. This is a problem of an entirely different order from teaching how to use AI.

Second, consensus comes before tools. The lesson of Korea's 1.4 trillion won experiment is not a failure of technology but an absence of consensus. If we decide 'which tools to teach with' before reaching social consensus on 'what to teach', the same failure will repeat. Trial and error will continue for more than a decade, but if we give up on the process, the dystopia of a humanity that does not think becomes reality.

Third, I believe a Renaissance comes when the conditions are created. It may not be a perfect analogy, but among the many factors behind the flowering of the Renaissance, including the growth of urban economies, the spread of humanism, and technological innovations like the printing press, what was decisive was that patrons like the Medici family created an environment in which creative people could devote themselves to their work free from worrying about making a living. Likewise, if the foundations of an education that lets our children question and think creatively are laid, and the social base to support it is in place, a new Renaissance in the AI era is not impossible.

We have a responsibility to give our children opportunities to think, continually. A society that turns away from that responsibility becomes dependent on platform companies, and that dependence is irreversible.

In the next installment, I will cover the utopia and dystopia AI creates in military and security .