## The question is no longer whether children will meet AI
A child does not need to open a chatbot to encounter artificial intelligence. AI already sits behind search results, recommendations, translation, image filters, writing suggestions, game opponents, moderation systems, adaptive learning platforms and the feeds that decide what appears next. For today’s pupils, AI is not a specialist topic waiting in a future computing lesson. It is becoming part of the ordinary information environment.
That changes the educational question. “Should students use AI?” is too blunt. A better question is: **what should a learner be able to understand, do and judge at each stage of development?**
An eight-year-old, a twelve-year-old and a sixteen-year-old should not receive the same AI lesson with different vocabulary. They are developing different capacities for abstraction, independence, source judgement, risk awareness and ethical reasoning. The goal is therefore not a single checklist called “AI literacy”. It is a progression.
The OECD and European Commission’s 2026 AI Literacy Framework for Primary and Secondary Education treats AI literacy as a combination of knowledge, skills and attitudes. Learners need to understand AI, evaluate what it produces, and use it ethically and creatively. That is a useful destination. But schools and families still need an operating map: what comes first, what must wait, what should become independent, and how can we tell whether a child is actually becoming more capable rather than merely becoming more fluent at prompting a machine?
The central principle of this article is simple: **AI literacy should mature from recognition, to verification, to judgement.** Young learners first need to know that AI output is made rather than magically known. Older learners need routines for checking claims and preserving their own thinking. Adolescents need to reason about evidence, bias, incentives, authorship, privacy, uncertainty and when not to delegate a task at all.
That progression matters because AI can create a dangerous educational illusion. A student may produce a cleaner paragraph, a faster answer or a more polished presentation while learning less. The OECD’s 2026 Digital Education Outlook makes the distinction sharply: generative AI can support learning when it is guided by sound teaching, but task outsourcing can improve immediate performance without producing corresponding learning gains. The visible product can rise while the learner’s independent capability stays flat.
AI literacy, then, is not “knowing how to use ChatGPT”. It is learning to remain an intelligent agent when intelligent-looking tools are available.
## 1. Start with a better definition: capability after the tool is gone
A weak definition of AI literacy focuses on operation. Can the student write a prompt? Can the student get an answer? Can the student make an image? These are tool-use skills. They may be useful, but they are not enough.
A stronger definition asks what remains after the screen closes.
Can the learner explain what the system did? Can they notice when an answer is uncertain? Can they identify what evidence would be needed to trust a claim? Can they distinguish assistance from substitution? Can they say which parts of a final product are theirs? Can they recognize a privacy risk before uploading personal information? Can they compare an AI answer with a textbook, data set, primary source or trusted expert? Can they reject an attractive answer that does not survive checking?
This gives us a practical test for any AI activity:
**Before:** What human capability is the student supposed to develop?
**During:** Which parts of the thinking may be supported by AI, and which parts must remain with the learner?
**After:** Can the student still perform the important part without the AI?
The third question is the decisive one. If the learner can only complete the task when the machine is present, we should be careful about calling the result learning. Sometimes assisted performance is exactly the point: translation support, accessibility, brainstorming, coding help or rapid feedback can remove unnecessary barriers. But when the instructional target is independent reasoning, writing, calculation, interpretation or recall, the final test must eventually remove the support.
This is why age-banding helps. Independence should grow with the learner. The amount of adult structure can fall as the learner’s verification, self-monitoring and ethical judgement become stronger.
## 2. At about age 8: teach that AI is a system, not a person who “knows”
For younger primary pupils, the first job is conceptual hygiene. Children naturally use social language with machines. A chatbot says “I think”, “I’m sorry” or “I can help”, and a young user can easily infer a mind behind the words. The system may feel like a very knowledgeable person who happens to live in a device.
The first lesson should gently break that illusion without making the technology mysterious. An eight-year-old does not need transformer architecture. They need a usable model:
**AI systems find patterns in data and use those patterns to produce outputs. The output can be useful, surprising or wrong.**
That sentence does several jobs. It explains why the machine can be impressive without being infallible. It also creates room for human checking.
At this age, useful lessons are concrete. Ask the class to compare three outputs: a correct answer, a partly correct answer and a confidently wrong answer. Let pupils circle which statements can be checked. Give them two images and ask whether an AI-generated picture represents a real event or an invented scene. Show that a fluent sentence can still contain a false fact. The aim is not to catch the system “failing” for entertainment. It is to build the habit that appearance and truth are different properties.
For younger learners, three rules are powerful:
1. **Do not share private information.** Names, addresses, passwords, identifying photographs, school details and personal family information need adult rules, not improvisation.
2. **Check important answers with a trusted source or adult.** AI can suggest; it does not automatically settle a question.
3. **Do your thinking before asking for the finished answer.** A child should first try, predict, draw, explain or choose. The AI can then become a comparison partner rather than a replacement thinker.
Notice what is absent: sophisticated prompt engineering. At eight, “better prompting” is not the educational centre of gravity. The learner’s most important development is epistemic: understanding that generated language is not the same thing as guaranteed knowledge.
## 3. The age-eight classroom: build the stop–ask–check reflex
Young learners benefit from short routines more than long policies. One useful routine is **Stop – Ask – Check**.
**Stop:** Is this an AI-generated answer, image or suggestion?
**Ask:** What is it claiming? Is the claim important?
**Check:** What could we use to verify it?
The check might be a teacher, a library book, a class text, a labelled diagram, a calculator, an atlas, a trusted website, a measurement or an observation. The exact source matters less than the move from passive acceptance to active verification.
A science example makes this clear. Suppose an AI assistant says that all insects have eight legs. A pupil who has learned only “AI sometimes makes mistakes” may simply become suspicious. That is not enough. A pupil with a verification routine asks: what do our science resources say? What can we observe about an ant or beetle? Which animal group typically has eight legs? The error becomes a route into classification and evidence.
In English, an AI tool might produce a polished continuation to a story. The question is not whether the writing is “good”. Ask the child which clues from the original story the continuation respects, which it ignores, and which words the child would keep or change. The learner stays in the author’s chair.
In mathematics, AI might solve a word problem correctly. The pupil still has to explain what each number represents and why the operation fits. A correct answer without an understandable route is not yet evidence that the pupil has learned the mathematics.
The important design feature is that AI use is attached to an existing learning target. We are not inventing artificial “AI time”. We are using ordinary subjects to teach the new judgement skills that AI now makes necessary.
## 4. At about age 12: move from “AI can be wrong” to “how do I verify this?”
Around the transition into secondary education, learners can handle a more demanding question. It is no longer enough to know that AI can make mistakes. They should start classifying *types* of risk.
A useful middle-school map includes at least six:
– **Factual error:** the answer contains a false claim.
– **Missing context:** the answer is technically true but incomplete in a way that changes meaning.
– **Weak evidence:** the answer makes a claim without showing a reliable basis.
– **Bias or framing:** the wording favours one perspective, stereotype or assumption.
– **Fabricated reference:** the system provides a source that does not exist or does not support the claim.
– **Over-delegation:** the answer may be fine, but the student has handed over the very thinking the lesson was meant to develop.
This is the age where verification should become explicit and teachable. “Check your work” is too vague. Students need a sequence.
For example: **Claim – Source – Match – Decision.**
**Claim:** What exactly is the AI saying? Rewrite it as a checkable statement.
**Source:** What independent source is suitable for checking it?
**Match:** Does the source actually support the same claim, or only something nearby?
**Decision:** Accept, revise, reject or leave uncertain.
That last option matters. Students should learn that “I cannot verify this yet” is an intelligent outcome. AI systems often encourage immediate completion because they always produce something. Education should protect the right to remain uncertain when evidence is insufficient.
At twelve, learners can also begin to see that sources have different jobs. A dictionary can confirm a definition but not settle a controversial historical interpretation. A government statistics portal can provide population data but may not answer a causal question. A peer-reviewed study can offer evidence about a specific sample without proving that an effect holds everywhere. AI literacy grows into source literacy.
## 5. Age twelve is where prompting should become purposeful, not performative
Prompting is often marketed as the central AI skill. It is certainly useful to specify a task clearly, provide constraints and ask for the form of answer you need. But a school can easily overteach prompt tricks and underteach judgement.
A better progression treats prompting as **task specification**.
Students should learn to state:
– the problem they are trying to solve;
– the information the tool may use;
– the constraints it must respect;
– the form of output that will be useful;
– what the system should flag as uncertain;
– what the student will verify independently.
This is more durable than memorising fashionable prompt formulas. Tools will change. The ability to define a problem, set constraints and inspect an output will remain useful.
Consider a Secondary 1 history learner. A weak prompt is: “Tell me about the causes of World War I.” A stronger educational task is: “Generate a table of four commonly discussed causes. For each, state one claim that needs verification and suggest what kind of historical source could test it. Do not provide invented quotations.” The second prompt does not merely seek content; it designs a checking workflow.
In English, instead of “improve my essay”, a student could ask the AI to identify sentences where the causal link is unclear, without rewriting them. The learner then repairs the reasoning. In science, the tool can generate plausible alternative explanations for an observation, and the student decides what evidence would distinguish them. In mathematics, AI can generate a similar problem after the student has solved one, then the student compares structures rather than copying steps.
Good use keeps the learner doing the intellectual move that the curriculum values.
## 6. At about age 16: teach delegation as a decision, not a default
Older secondary students are approaching adult digital life. They will meet AI in coursework, employment, creative work, research, administration and personal decision-making. At this stage, AI literacy should become a question of governance: **what should I delegate, under what conditions, and what responsibility remains mine?**
A useful distinction is between four types of task.
**Low-stakes mechanical tasks:** reformatting a table, generating practice items, converting notes into flashcard candidates. Delegation may be efficient if the output is checked.
**Supportive cognitive tasks:** brainstorming alternatives, receiving critique, asking for counterarguments, comparing explanations. AI can widen the field while the learner retains judgement.
**Core learning tasks:** deriving a proof, writing an argument, analysing a source, solving an unfamiliar problem. If these are outsourced too early, the final product can conceal missing capability.
**High-stakes judgement tasks:** decisions involving safety, health, legal consequences, major financial choices, safeguarding or serious accusations. AI output here requires strong external authority and should not be treated as the decision-maker.
Sixteen-year-olds can understand that efficiency is not always the objective. A gym machine that lifts the weight for you makes the movement easier but changes the training effect. Likewise, an AI system that produces the argument, code or explanation may make the assignment easier while removing the mental work that would have built capability.
The mature question is not “Can AI do this?” It is “**What happens to me if AI does this part for me?**”
## 7. The generation line: who produced the reasoning?
One of the most useful concepts for older students is the **generation line**: the boundary between thinking produced by the learner and thinking generated externally.
The line does not need to sit in one place for every task. A student learning a new topic may receive more scaffolding. A student demonstrating mastery should cross less of the core work to the machine.
A practical classroom protocol is:
**Attempt → Assist → Audit → Reperform.**
**Attempt:** The student makes a genuine first attempt. It can be incomplete.
**Assist:** AI is used for a defined purpose: feedback, examples, alternatives, explanation or checking.
**Audit:** The student inspects the output, verifies claims and records what changed.
**Reperform:** The student completes a fresh version or a related task without the same assistance.
Reperformance is the part most likely to be skipped. Yet it is the strongest evidence that the tool helped learning rather than merely improving a product.
Imagine a student uses AI to diagnose weaknesses in an argumentative paragraph. The paragraph becomes much better. That is useful, but it does not prove the learner has internalised the lesson. Give the student a new paragraph problem a day later. Can they identify the missing warrant, weak evidence or vague qualifier independently? If yes, the assistance may have transferred into capability. If not, the polished original was mainly assisted performance.
This is a fairer way to evaluate AI use than trying to detect whether text “looks AI-generated”. The educational question is not only authorship. It is capability.
## 8. Verification should become a subject-independent discipline
By sixteen, students should be able to verify an AI answer with more than one method. Different claims require different checks.
For numerical claims, recomputation and original datasets may be appropriate. For scientific claims, students may need a reputable review, textbook or primary study. For historical claims, they can compare dates, sources, provenance and interpretations. For current affairs, they should check publication date, named evidence and multiple credible outlets. For quotations, they should find the original text. For code, they should test outputs and edge cases rather than assuming clean syntax means correct behaviour.
This can be taught with a **verification ladder**:
1. **Internal sense check:** Does the answer contradict something already known? Are units, dates or causal links suspicious?
2. **Direct check:** Can I calculate, measure, run, inspect or reproduce it myself?
3. **Source check:** Can I find an authoritative independent source?
4. **Cross-source check:** Do multiple credible sources agree, and are they independent?
5. **Boundary check:** What exactly remains uncertain, context-dependent or contested?
The fifth step prevents verification from becoming a ritual of finding one webpage that agrees. Mature information literacy includes knowing the limits of a conclusion.
AI raises the value of this skill because fluent answers are cheap. When words become easy to generate, the scarce skill shifts toward selecting, checking and judging them.
## 9. Teach provenance: where did this information come from?
Students also need a stronger sense of provenance: the origin and chain of information.
A chatbot response is often a convenient surface. It may summarise ideas originating in textbooks, websites, papers, databases or patterns in its training data. The surface itself is not automatically the best evidence. When a task requires academic reliability, students should move toward the underlying source.
A useful habit is: **AI can help me find the question; evidence has to help me settle it.**
For research tasks, students can ask AI for search terms, alternative hypotheses or a reading plan. But claims in the final work should be traceable to sources the student has actually opened and checked. If a citation cannot be found, it cannot be treated as evidence. If a source exists but does not say what the AI claimed, the mismatch must be corrected.
This is especially important because fabricated citations are uniquely persuasive. They have the appearance of scholarship: author names, journal titles, publication years and plausible article titles. A student who treats citation formatting as proof can be misled. The verification skill is not “does this look like a reference?” but “can I retrieve it, identify it and confirm that it supports the claim?”
By upper secondary, this should be routine, not a special warning delivered before a research project.
## 10. Bias is not only an AI problem; AI makes framing easier to overlook
Students often hear that AI can be biased. The phrase is correct but too abstract unless learners can see what bias looks like in practice.
Bias can enter through data, labels, sampling, historical patterns, system objectives, safety filters, prompt wording or the human interpretation of an output. The educational move is to ask whose perspective is represented, what comparison group is missing, and what assumption the wording quietly makes.
A geography response might describe one country mainly through poverty indicators while describing another through innovation. Both sets of facts could be true, yet the framing produces unequal pictures. A careers tool might associate leadership more often with certain demographic patterns. A literature explanation may overstate a dominant interpretation and omit plausible alternatives. A language model may answer a culturally specific question using assumptions drawn from another context.
Older students should practise reframing prompts deliberately: “What assumptions are built into this question?” “What would someone from another context challenge?” “Which groups might be underrepresented in the evidence?”
The aim is not to make students distrust all AI output. It is to make them notice that any information system can frame reality, and generated language can hide that framing behind smooth prose.
## 11. Ethics should move beyond plagiarism rules
If AI ethics is reduced to “do not cheat”, students learn only compliance. Ethical AI literacy is broader.
Learners should consider privacy, consent, attribution, fairness, misinformation, intellectual ownership, labour, environmental cost, accessibility and the consequences of automated decisions. Not every issue needs equal depth at every age, but the map should widen over time.
At eight, ethics may mean: do not upload someone else’s photo without permission; do not use AI to make a mean image of a classmate; tell an adult if a system asks for personal information.
At twelve, it can include: disclose meaningful AI assistance; do not fabricate evidence; understand that an AI image of a real person can cause harm even if labelled as a joke; distinguish private data from safe task information.
At sixteen, students can debate harder cases. If AI restructures a paragraph but the ideas are yours, what should be disclosed? If a tool recommends a candidate for a job interview, who is accountable for bias? If an artist’s work helped train a model, what questions arise about consent and compensation? If a model can generate persuasive political content at scale, what happens to public trust?
These questions help students see AI as part of society rather than a clever school utility.
## 12. Do not confuse restriction with literacy
Schools need rules. Some tasks should be completed without AI because independent performance is the target. Some contexts involve privacy or safeguarding constraints. Examinations may prohibit assistance. Younger children may need supervised access.
But prohibition alone does not create literacy. A student can follow “no AI” rules in school and still use AI poorly everywhere else.
The stronger approach combines boundaries with instruction. A school can say: “No AI for this first draft because we need to see your unaided writing,” and later say, “Now use AI to generate two critiques; decide which critique is valid and revise only where you agree.” The rule and the learning purpose are visible.
Likewise, a mathematics teacher may prohibit AI during a diagnostic because the teacher needs a clean signal of understanding, then allow it during error analysis to generate alternative solution routes for comparison.
The key is explainability. Students should know why access changes. Arbitrary rules invite workaround behaviour. Purpose-linked rules teach decision-making.
## 13. Build an AI literacy progression across subjects, not a one-off digital citizenship lesson
AI literacy is too important to live only inside computing or a yearly assembly. Different subjects expose different forms of judgement.
English can teach authorship, rhetoric, source use, tone and revision. Mathematics can teach verification, constraints, model assumptions and the difference between an answer and a proof. Science can teach evidence, uncertainty, reproducibility and competing explanations. Humanities can teach provenance, perspective, historical context and framing. Art can teach originality, transformation, attribution and creative intent.
A school progression might therefore specify capabilities rather than software:
**Primary:** recognise AI-generated output; protect personal information; try first; verify simple claims; distinguish real from invented evidence.
**Lower secondary:** classify errors; check sources; compare perspectives; use AI for bounded support; disclose assistance; identify when a task should remain human-generated.
**Upper secondary:** design verification workflows; manage provenance; evaluate bias and uncertainty; decide what to delegate; preserve independent capability; explain ethical trade-offs; audit high-stakes claims.
This is much more robust than teaching one vendor interface. Interfaces will change. The capability stack survives.
## 14. Assessment must test judgement, not just recall of AI vocabulary
A student can memorise definitions of “hallucination”, “bias” and “prompt” without being able to use them. Assessment should therefore include decisions.
Give students an AI answer with three embedded problems and ask them to identify which claim is false, which is unverifiable and which is acceptable. Ask them to design a checking plan. Provide two sources and ask which one is better for a particular claim. Give them a task and ask which parts could be delegated without damaging the learning objective. Ask them to disclose AI assistance accurately. Let them compare an AI-generated argument with a human-written one and judge evidence rather than guessing authorship.
For older learners, an excellent assessment is a **tool-removal transfer task**. Let students use AI during guided practice, then give them a fresh problem where the relevant reasoning must be performed unaided. If performance transfers, the tool may have supported learning. If it collapses, the activity may have trained dependence.
This also protects teachers from being impressed by surface polish. The cleanest product is not automatically the best evidence of education.
## 15. Families need simple language for AI at home
Parents do not need to become machine-learning specialists. They need questions that keep the child thinking.
Instead of “Did you use AI?”, try:
– What did you try before you asked it?
– Which part of the answer did you check?
– What did the AI change in your thinking?
– What did you keep because you judged it better?
– What would you be able to do without the tool now?
– Did you share anything private?
– If the answer matters, where else did you verify it?
These questions shift the conversation away from policing and toward intellectual responsibility.
Families can also create risk tiers. Brainstorming a birthday-card idea is low stakes. Checking a homework fact is medium stakes and needs verification. Medical, legal, financial or safeguarding decisions are high stakes and require qualified human sources. Children can understand that not all questions deserve the same trust threshold.
## 16. A practical developmental map
Here is a compact version schools can adapt.
### Around age 8: Recognise and check
The learner should increasingly be able to:
– identify when an output may be AI-generated;
– explain that AI uses patterns and can be wrong;
– protect personal information;
– attempt a task before requesting a finished answer;
– verify simple factual claims with a trusted source;
– distinguish a real source from a made-up claim;
– tell an adult when something generated is confusing, upsetting or unsafe.
### Around age 12: Verify and compare
The learner should increasingly be able to:
– separate factual error, missing context, weak evidence and bias;
– turn a generated answer into checkable claims;
– select an appropriate independent source;
– compare multiple answers or perspectives;
– use AI for bounded help instead of full substitution;
– disclose meaningful AI assistance;
– explain why some school tasks need to remain unaided;
– recognise that fluent writing is not proof of truth.
### Around age 16: Judge and govern
The learner should increasingly be able to:
– decide which cognitive work to delegate and which to retain;
– verify sources, quotations, data and citations independently;
– reason about uncertainty and contested evidence;
– test generated code, calculations or models rather than trusting appearance;
– analyse bias, framing and stakeholder impact;
– preserve authorship and accountability;
– evaluate ethical trade-offs;
– reperform important skills without the tool;
– explain the limits of an AI-supported conclusion.
The ages are not hard gates. Learners develop unevenly, and schools will adapt the progression to context. What matters is the direction: from **recognition**, to **verification**, to **judgement**.
## 17. The deepest AI literacy skill is knowing when not to ask
A mature learner does not use AI at every opportunity. They know that some forms of difficulty are productive.
If you are learning to write, the struggle to find the right sentence is part of the training. If you are learning mathematics, the delay before selecting a method is part of becoming better at method selection. If you are learning to read a demanding text, the work of making connections is part of comprehension. If you are studying for an examination, retrieving an answer from memory is different from recognising a generated explanation.
AI can support each of these activities, but timing matters. Assistance delivered before an attempt may remove the very cognitive event that would have strengthened the learner.
This is why the developmental map ends with restraint. The most capable sixteen-year-old is not the student with the longest prompt library. It is the student who can look at a task and say:
**I know what the machine could do here. I also know what I need to do myself.**
That is not anti-technology. It is agency.
## 18. Three lesson designs that show the progression in practice
The same topic can be used to teach very different levels of AI literacy. Consider a simple question: **Why do some cities feel hotter than nearby rural areas?**
For an eight-year-old, the teacher might ask an AI system for a short explanation and place it beside a child-friendly science source. Pupils identify two claims, find where the trusted source supports or contradicts them, and mark one sentence that sounds confident but still needs checking. The learning objective is recognition and verification, not urban climate science in full depth. The child experiences the idea that an answer can be readable without automatically being reliable.
For a twelve-year-old, the teacher can ask students to generate two explanations of the urban heat island effect: one accurate and one containing a plausible misconception. Students must identify the questionable claim, choose an appropriate source, and record a Claim–Source–Match–Decision trail. Then they rewrite the explanation using only information they have verified. The important skill is not detecting an “AI voice”. It is checking a proposition.
For a sixteen-year-old, the task can become a mini research audit. Students ask AI to propose factors affecting urban heat, then classify each factor as a hypothesis, established mechanism or claim requiring evidence. They locate sources, compare whether the evidence transfers across climates, and note uncertainties. Finally, they decide which parts of the workflow benefited from AI and which parts would have undermined the learning if delegated.
The progression is visible. The youngest learner asks, “Can I trust this?” The middle learner asks, “How do I verify this?” The older learner asks, “What is the strength and boundary of the evidence, and what responsibility remains mine?”
A second example can use writing. At eight, AI can provide two endings to a story and pupils decide which better fits the characters and clues. At twelve, AI can identify places where a paragraph lacks evidence without rewriting it. At sixteen, students can ask for counterarguments, then verify whether the counterarguments are fair representations before revising their own position.
A third example can use mathematics. At eight, students compare an AI answer with counters or drawings. At twelve, they inspect two solution routes and find the first invalid step. At sixteen, they can ask AI to generate edge cases that test whether a general claim holds, then prove or refute the claim independently.
These lesson designs share a rule: the tool changes, but the learner keeps the decisive intellectual job.
## Conclusion: teach children to stay in charge of the learning
AI literacy belongs beside reading, writing, mathematics, science and information literacy because it changes the conditions under which all of them are practised. But the educational response should not be panic, novelty chasing or prompt theatre.
At eight, children need to recognise that generated output is not guaranteed knowledge. At twelve, they need disciplined verification and bounded use. At sixteen, they need judgement about delegation, provenance, uncertainty, ethics and independent capability.
The destination is not a child who can make AI produce impressive things. The destination is a young person who can use powerful tools without confusing their output with truth, their fluency with understanding, or their assistance with personal mastery.
When AI gets easier to use, human judgement becomes more important, not less.
—
## Related eduKateSG reading
– **How Education Works | AI Literacy Education — How Learners Understand, Evaluate, Use and Shape Artificial Intelligence**
How Education Works | AI Literacy Education — How Learners Understand, Evaluate, Use and Shape Artificial Intelligence
– **How to Think Properly | Use AI Without Handing Over the Final Judgement**
How to Think Properly | Use AI Without Handing Over the Final Judgement
– **How Studying Works | The Locus of Generation — When AI Produces the Explanation, Who Did the Learning Work?**
How Studying Works | The Locus of Generation — When AI Produces the Explanation, Who Did the Learning Work?
– **How Studying Works | Verification Burden — Why Easier Answers Make Independent Judgment More Valuable**
How Studying Works | Verification Burden — Why Easier Answers Make Independent Judgment More Valuable
## Sources and further reading
– OECD & European Commission (18 June 2026), *Empowering Learners for the Age of AI: An AI Literacy Framework for Primary and Secondary Education*.
https://www.oecd.org/en/publications/empowering-learners-for-the-age-of-ai_65cd27d4-en.html
– OECD (19 January 2026), *OECD Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education*.
https://www.oecd.org/en/publications/oecd-digital-education-outlook-2026_062a7394-en.html
– Harvard Graduate School of Education (17 December 2025), *What We Learned: Our 2025 Era* — reports that AI content dominated its most-visited education stories in 2025.
https://www.gse.harvard.edu/ideas/news/25/12/what-we-learned-our-2025-era
– American Psychological Association (10 December 2025), *How to help your students use AI without losing the learning*.
https://www.apa.org/ed/precollege/psychology-teacher-network/introductory-psychology/learning-artificial-intelligence
