Education has two AI problems, not one
Generative AI can help a learner perform a task and can help the learner acquire the underlying skill. Those outcomes are not the same. A polished essay produced with extensive assistance may improve today’s submission while weakening the practice needed to write independently. Conversely, a tutor that asks questions, diagnoses misconceptions, and gradually removes support can improve learning. The educational question is therefore not simply whether students “use AI,” but what cognitive work the student still performs.
Current classroom effects concern generative AI, not AGI. Schools already face fluent text generation, translation, feedback, coding help, synthetic media, and automated administration. Claims that AGI will make schools or degrees obsolete are projections. No current evidence can establish how institutions would respond to a system with reliable, broadly superhuman teaching and research abilities.
What is observed as of 2026
Student use has become mainstream in many participating education systems. OECD’s PISA 2025 results reported that 46% of students in OECD countries used AI chatbots weekly or more to help them learn. Students using AI for particular schoolwork tasks such as summarizing, drafting, and preliminary research had lower science scores on average after adjustment for socioeconomic status, while general weekly users performed similarly to non-users. This is an association, not proof that AI caused lower scores; struggling students may use it differently or more often. Students with both regular use and school opportunities to develop AI literacy tended to score slightly higher, highlighting the importance of instruction and inequality of access (OECD PISA 2025, September 2026).
Teacher readiness lags student adoption. Stanford’s 2025 AI Index reported that 81% of surveyed U.S. computer-science teachers thought AI should be included in foundational learning, while fewer than half felt equipped to teach it. That is a U.S. teacher finding, not a global measure, but it illustrates why policy documents alone do not create competent classroom practice (Stanford HAI, 2025).
Cheating is real but hard to measure. Surveys depend on self-report and definitions differ: prohibited ghostwriting, permitted brainstorming, grammar correction, and accessibility support may all be counted as “AI use.” AI-text detectors add another problem. Their error rates vary by model, language, editing, and student population; a probabilistic detector should never be the sole basis for punishment. Institutions need evidence-based academic-integrity procedures with notice, an opportunity to respond, and attention to false positives.
What early intervention evidence suggests
The strongest finding is conditional: AI support works better when embedded in sound pedagogy. The OECD’s 2026 Digital Education Outlook concludes that generative systems may provide personalized feedback and tutoring, but distinguishes AI-augmented performance from durable skill development and calls for pedagogically grounded use (OECD, 2026). A tool’s vendor has an interest in reporting engagement and short-term gains, so independent randomized studies, delayed tests, and transfer to new problems deserve more weight than testimonials.
Useful tutoring behavior includes asking a learner to attempt the problem first, giving hints rather than answers, requiring explanation, checking sources, revisiting errors, and fading assistance. Harmful patterns include completing the entire task immediately, inventing citations, praising incorrect reasoning, or adapting so completely to a learner’s preferences that productive difficulty disappears.
Human teachers remain essential for goals the model cannot reliably own: choosing curricula, understanding family and cultural context, managing a classroom, protecting children, motivating sustained effort, recognizing distress, and accepting professional accountability. AI can expand feedback capacity without making those responsibilities disappear.
Assessment must measure the intended capability
Traditional take-home essays cannot carry the same evidentiary weight when high-quality assistance is ubiquitous. The response should not be surveillance or a return to memory-only examinations. Programs can combine:
- supervised, no-AI demonstrations of foundational knowledge;
- open-AI assignments that require logs, source checking, and critique;
- oral defense or live explanation;
- drafts and process artifacts;
- projects tied to local data or direct observation;
- collaborative work with explicit individual accountability.
Each mode answers a different question. A no-tool task tests unaided fluency. An open-tool task tests judgment and orchestration. An oral defense tests whether the student understands the submitted work. Accessibility accommodations must remain available, and assessment changes should not penalize students for using assistive technology.
Accuracy, sources, and intellectual dependence
Chatbots can produce correct explanations and fabricated claims in the same confident register. Students should learn lateral reading: leave the generated answer, inspect primary sources, identify the author and date, trace statistics to their method, and seek contradictory evidence. Citing a chatbot is not a substitute for citing the underlying source.
AI literacy also includes recognizing uncertainty, data privacy, model bias, synthetic media, and commercial incentives. A school account may have different retention terms from a consumer account. Students should not enter identifiable peer information, protected education records, unpublished research, or intimate disclosures without an approved system and clear policy.
There is a deeper risk of intellectual dependence. If a learner delegates topic selection, search, synthesis, drafting, and revision, the final product may exceed the learner’s capability while leaving little knowledge behind. Teachers can counter this by specifying which steps may be assisted and requiring reflection on where the model was wrong or unhelpful.
Equity cuts in both directions
AI can provide translation, text simplification, speech support, and individualized practice. Those benefits may be especially important for students with disabilities or limited access to tutoring. Yet paid models, fast devices, connectivity, and teacher expertise are unequally distributed. OECD found AI-literacy opportunities more common among socioeconomically advantaged students. A nominally universal technology can widen gaps if affluent students receive guided, high-quality use while others receive unrestricted answer generation or no access.
Equity policy should therefore provide accessible approved tools, teacher development, alternatives for students who opt out, and evaluation across language and disability groups. It should also preserve non-AI routes so education does not become contingent on sharing student data with a private provider.
Credentials and the labor market
Employers may place less value on credentials based mainly on take-home production that AI can imitate. That does not mean degrees become worthless. Credentials also signal sustained effort, supervised practice, laboratory or clinical competence, peer collaboration, and institutional verification. Their value will depend on whether schools can demonstrate what graduates can do independently and with tools.
Curricula should preserve foundational knowledge because judgment requires something to judge with. At the same time, students need practice decomposing problems, evaluating outputs, working across disciplines, communicating with people, and taking responsibility for consequences. “Prompt engineering” alone is too tool-specific to be an educational philosophy.
A practical policy for families and schools
Adults should ask: What learning objective is this assignment testing? Which assistance is allowed? Must AI use be disclosed? What data leaves the institution? How can a student challenge an accusation? What happens when the tool is unavailable?
Families can ask children to show their process, explain answers in their own words, verify one important claim together, and maintain regular unaided reading, writing, mathematics, and conversation. Blanket prohibition can drive use underground; total permissiveness can replace practice with performance.
The best-supported position in September 2026 is structured integration with continued evaluation. Current evidence does not justify claiming that generative AI universally improves or damages learning. It shows that effects depend on task, pedagogy, intensity, prior skill, and access. AGI could change the economics and reach of education dramatically, but that remains scenario analysis rather than a result schools can observe today.