The assignment is not the learning
An AI system can produce a correct essay, solve an equation, explain a historical event, and adapt an example to a student’s interests. None of those outputs proves that the student can reason without the system, detect its mistake, apply the idea in a new setting, or decide which question is worth asking. Education after transformative AI begins with that distinction.
The near-term evidence is already cautionary. The OECD Digital Education Outlook 2026 synthesizes emerging studies in which pedagogically designed AI supported learning, while general-purpose chatbots sometimes improved immediate task performance without durable gains. In one cited high-school mathematics field experiment, students using a standard GPT-4 interface performed 17% worse than controls on a later unassisted test, while a tutor with learning safeguards avoided that penalty. That result concerns one task and implementation; it is not evidence that every chatbot reduces learning. Tool-assisted output and acquired capability must be measured separately.
This becomes more important as systems improve. When an answer is cheap, the scarce educational work is building models of the world, practicing judgment, developing motivation, learning with other people, and becoming capable of checking a powerful assistant rather than merely operating it.
Education still has at least five jobs
Schools are not only information-delivery systems. They help children acquire foundational skills, encounter domains they would not choose alone, learn social expectations, form relationships, and receive credentials that other institutions recognize. They also provide meals, safety, disability services, counseling, sports, arts, and childcare that enables adults to work.
An excellent tutor can change the instructional part without replacing the institution. A post-AGI education policy that counts only explanations per dollar will miss most of what families and communities need from schools.
The five jobs can be stated plainly:
- Capability: build knowledge and skills a person can use independently.
- Judgment: practice deciding what evidence to trust and what values should guide action.
- Development: support physical, emotional, social, and moral growth.
- Membership: give young people real places in communities with obligations to others.
- Certification: provide credible evidence of competence without turning surveillance into the curriculum.
AI can assist each job. It should not be allowed to silently redefine them around what the model can measure.
Personalized tutoring: real promise, specific limits
One-to-one tutoring has historically been expensive. AI can make explanations, practice questions, translation, feedback, and accessibility support available at extraordinary scale. A student can ask the same question ten ways without embarrassment. A teacher can generate several examples and spend more time diagnosing why a learner is stuck.
The promise depends on implementation. The UNESCO guidance for generative AI in education is non-binding policy guidance emphasizing human agency, privacy, age appropriateness, inclusion, and pedagogical validation. A fluent system can fabricate facts, adapt to the wrong learning objective, expose sensitive student data, or optimize for continued use rather than eventual independence.
A useful tutoring system should make itself progressively less necessary. It asks the learner to retrieve knowledge before supplying it, varies problems rather than repeating a template, shows uncertainty, encourages outside verification, and periodically tests performance without assistance. Success is not the number of pleasant conversations. It is what the learner can later do.
Teachers remain essential because instruction is relational and contextual. A teacher sees frustration, family change, peer dynamics, avoidance, unusual talent, and a pattern of misunderstanding across weeks. Even a technically superior explanation may fail if nobody helps a child tolerate difficulty or believe that effort will matter.
Foundational knowledge becomes more important, not less
People sometimes argue that students no longer need facts because a model can retrieve them. That confuses access with understanding. Background knowledge lets a person notice an implausible answer, connect ideas, form a useful query, and hold enough context to reason. Without it, verification collapses into asking another model and choosing the response that sounds best.
Foundational literacy, numeracy, scientific reasoning, civics, history, statistics, media literacy, and basic computing remain protective infrastructure. Memorization should not dominate education, but retrieval practice and internal knowledge are part of building a mind that can operate when tools are absent, manipulated, or wrong.
The curriculum also needs model literacy: how training data differ from evidence, why generated citations fail, what uncertainty means, how automated rankings shape attention, and when private information should never enter a system. This belongs inside existing subjects rather than as one disposable technology lesson.
Assessment must move closer to performance
Take-home essays and routine problem sets become weak evidence of individual ability when a model can complete them. Responding with universal surveillance or automated cheating detection would create its own harms and still produce false accusations.
Assessment can instead use several forms of evidence:
- supervised writing and problem solving;
- oral explanation and defense of choices;
- portfolios showing revisions and process;
- practical demonstrations and projects;
- collaborative work with individual reflection;
- explicit AI-assisted assignments that require checking, comparing, or improving model output;
- periodic tool-free assessment of capabilities that must remain independent.
No single method is perfectly authentic or equitable. A portfolio can receive outside help; oral exams can disadvantage some students; timed work can measure anxiety. A mixed system is harder to game and gives learners more than one way to demonstrate competence.
Credentials may also shift. Employers and institutions could rely more on demonstrated work, apprenticeships, licensed examinations, and continuously updated records of competence. That can open opportunities, but it can also become continuous behavioral scoring. Educational records should remain purpose-limited, correctable, and under rules that prevent a childhood model profile from following a person indefinitely.
Growing up requires encounters that are not optimized
An AI companion can be endlessly patient, interested, and responsive. Human peers are none of those things consistently. Children learn negotiation, repair, empathy, boundaries, humor, leadership, and tolerance for difference through relationships they cannot fully control.
A system optimized to retain attention may remove exactly the friction through which social development occurs. It may affirm a child when a responsible adult would challenge them, or imitate intimacy without carrying reciprocal needs. Schools and families should preserve device-free play, team activity, intergenerational relationships, unstructured time, and real responsibility for other people.
This is not an argument for banning AI from childhood. Translation, accessibility, creative tools, tutoring, and communication support can expand participation. The boundary is that simulated relationship should not quietly replace human attachment, professional mental-health care, or the difficult practice of living with peers.
Normative priority: teachers need governance power, not only training
“Train teachers to use AI” is incomplete when teachers have no say over procurement, data use, workload measurement, or whether a system’s output can override professional judgment. Educators, students, parents, disability specialists, and privacy staff should participate before adoption.
Contracts should state what student data are collected, whether conversations train models, how long records remain, which subcontractors receive them, and how schools can export or delete data. Systems should be evaluated across languages, disability needs, age groups, and local curricula. A school must retain an effective non-AI path for students who cannot safely or lawfully use the product.
AI should reduce administrative burden instead of generating new dashboards that intensify monitoring. Time saved on routine drafting should move toward feedback, planning, and relationships—not automatically justify larger classes or fewer support staff before learning outcomes are known.
Access is more than a free chatbot
Low-cost tutoring can broaden opportunity only if students have connectivity, devices, language support, quiet places to study, and adults who can help. Wealthier families may combine the same systems with human tutors, enrichment, safe peer networks, and recognized credentials. Providing software without complementary institutions can leave inequality intact.
Public libraries, schools, community colleges, and trusted civic organizations can provide access with privacy protection and human support. Public evaluation can test products against educational goals rather than engagement. Open educational materials can reduce dependence on one vendor and let communities inspect what is taught.
Internationally, language coverage and local knowledge matter. A system strong in English and wealthy-country curricula is not a universal tutor. Communities need the ability to build, translate, correct, and govern educational resources without surrendering student data or cultural authority.
The post-AGI graduate
Education should not try to beat machines at every task. It should prepare people to remain authors of goals, competent judges of evidence, responsible participants in institutions, and capable friends, family members, neighbors, and citizens.
A graduate in this world should be able to work with advanced systems and without them when necessary; distinguish an answer from an argument; recognize uncertainty and manipulation; learn a difficult domain; cooperate with people; protect private information; and take responsibility for a decision that affects others.
Those aims are older than AI. Transformative capability makes them more urgent because outsourcing cognition becomes easy long before anyone knows which human capacities will be difficult to rebuild.