A twelve-year-old asks an AI system to explain a historical event.
The answer is fluent. It includes dates, names and a causal story. One detail is wrong. Another is contested. A third is phrased with more certainty than the source record warrants.
The student has several possible tasks at once.
They need to know enough about AI to understand why a confident answer can still be wrong. They need enough information literacy to check the claim. They need enough ethical judgment to consider what should be disclosed if AI helped produce an assignment. They need enough creative competence to use the system without surrendering authorship. And they need enough civic imagination to ask what kind of AI-shaped society they are participating in.
That cluster of capacities is the problem the 2026 AI Literacy Framework for Primary and Secondary Education attempts to organize.
What the framework is
On 18 June 2026, the OECD and the European Commission published Empowering Learners for the Age of AI: An AI Literacy Framework for Primary and Secondary Education.
The report is 64 pages and carries DOI 10.1787/65cd27d4-en. It presents AI literacy as a combination of knowledge, skills and attitudes intended to help learners understand, evaluate and use AI responsibly and creatively.
The framework is designed as a common reference for primary and secondary education rather than as a single subject syllabus.
That distinction matters.
It is meant to support curriculum design, teaching, policy and learning outcomes across education systems. It is not presented as a standardized psychological test of AI literacy or as evidence that one curriculum model has already been shown to work better than another.
Four domains: engage, create, manage and shape
The final 2026 framework organizes AI literacy into four broad domains:
Engage with AI. Learners need to recognize when and how AI systems are involved, interrogate outputs, understand relevant capabilities and limits, and interact with AI systems with appropriate critical attention.
Create with AI. Learners need to use AI as part of generative and productive activity while retaining authorship, purpose, judgment and responsibility.
Manage AI. Learners need to make informed choices about when to use AI, how to configure or constrain its role, how to protect information and how to manage risks, dependence and error.
Shape AI. Learners need to understand that AI systems and their social uses are not inevitable facts of nature. They are designed, governed and contested human systems. Learners should be able to participate in decisions about how AI should be built and used.
The final wording matters because earlier draft materials used somewhat different labels. A responsible reference page should distinguish the final publication from preliminary versions rather than blending them.
AI literacy is not prompt literacy
One of the framework's strongest contributions is conceptual.
It prevents AI literacy from collapsing into the ability to write effective prompts.
Prompting can be useful. But a person who can reliably obtain polished output from a generative system may still be unable to judge whether the output is true, appropriate, lawful, fair or worth acting on.
Operational fluency is not the same as epistemic competence.
The distinction becomes especially important in schools because apparent productivity can disguise dependency. A student may complete more polished work while practicing less recall, less source evaluation or less original synthesis.
The educational question therefore cannot be simply: Can the learner use AI?
It must also be: Can the learner decide when AI should be used, inspect what it produces, identify what remains uncertain, preserve authorship and take responsibility for the consequences?
A competence framework is a map, not a causal study
The framework has policy authority, but policy authority and experimental evidence are different things.
This publication draws together research, expert consultation and educational priorities to define a coherent competence architecture.
It does not randomly assign schools to competing AI-literacy programs and measure long-term outcomes.
It therefore cannot establish that implementing the framework causes better critical thinking, safer AI use or improved learning.
Nor does it validate a single score that can rank one learner's AI literacy against another's.
This is not a defect. It is a category boundary.
Frameworks and trials answer different questions.
A framework asks what educational capabilities should be named, organized and made teachable.
An intervention study asks whether a particular instructional approach changes specified outcomes under specified conditions.
An assessment study asks whether those outcomes can be measured validly and reliably.
A mature education system needs all three.
Why the framework belongs inside discernment
AI literacy is partly technical, but many of its difficult problems are epistemic.
A generative model produces language that can look authoritative without bearing the ordinary traces by which readers judge expertise. It can blend strong sources with weak ones. It can provide citations that need verification. It can mirror a user's premise even when that premise is false.
Using such systems well therefore requires more than knowledge of model architecture.
It requires source criticism, uncertainty awareness, calibration, lateral verification, critical thinking and the ability to distinguish fluent output from warranted belief.
This is why the framework sits naturally beside the Institute's Discernment & Epistemic Agency hub.
The overlap does not mean that AI literacy and discernment are identical. AI literacy also includes technical, creative, ethical and governance dimensions. But discernment is one of the capacities that makes AI literacy consequential rather than cosmetic.
The framework makes responsibility visible
The domain of shaping AI is especially important.
Young people are often spoken about as future users of technology. That language can quietly imply that the technology arrives complete and their task is adaptation.
The framework offers a broader view.
AI systems embody choices about training data, goals, interfaces, access, moderation, labor, environmental cost, privacy and institutional power. Learners can be taught to recognize those choices and participate in civic debate around them.
This moves AI literacy from consumer competence toward agency.
It also prevents responsibility from being placed entirely on individual students. If schools adopt AI tools, institutions remain responsible for procurement, privacy, accessibility, teacher support, assessment design and safeguarding.
A literate learner cannot compensate for a badly governed system.
Developmental appropriateness remains a major implementation problem
A framework spanning primary and secondary education has to cover large developmental differences.
The epistemic demands appropriate for an eight-year-old are not those appropriate for a seventeen-year-old.
Young learners may begin with recognizing that computer-generated answers can be mistaken and that important claims should be checked with a trusted adult or source.
Older learners may engage with model uncertainty, training data, representational harms, intellectual property, privacy, statistical evaluation and the politics of automated decision systems.
The framework provides an architecture, but implementation still requires developmental sequencing.
That sequencing should be tested rather than assumed.
Assessment is the next difficult layer
Once a competence enters a framework, institutions naturally want to assess it.
This creates risk.
AI literacy is multidimensional. A short quiz of AI vocabulary may measure factual knowledge while missing judgment in authentic situations. A performance task may capture practical competence but be difficult to score consistently. Self-report can capture attitudes while being vulnerable to confidence, social desirability and inaccurate self-assessment.
Assessment should therefore be aligned to the construct being claimed.
If the claim is about source verification, observe source verification.
If the claim is about calibrated trust, include situations in which trust should increase and situations in which it should decrease.
If the claim is about ethical reasoning, do not infer it from software fluency.
The more consequential the use of the assessment, the stronger the validity argument must be.
The Northern European significance is practical, not decorative
For Nordic and European education systems, this framework offers a shared vocabulary at a moment when schools are making rapid decisions about generative AI.
That common vocabulary can help ministries, municipalities, teacher education programs and schools distinguish curriculum goals from procurement choices and from assessment policy.
It can also support comparison across countries without assuming identical implementation.
Northern Europe has strong traditions of media literacy, democratic education and public-sector digitalization. Those traditions are relevant resources, but they are not guarantees of AI literacy.
High digital access can coexist with weak verification habits. Strong educational institutions can still adopt tools faster than evidence accumulates.
The framework is therefore useful precisely because it makes the desired capacities discussable before they are reduced to platform skills.
Knowledge, skills and attitudes should not be collapsed into one competency score
The framework deliberately includes more than factual knowledge.
A learner may know what a training dataset is but still trust outputs uncritically. Another may be skeptical of AI but unable to use it productively. A third may operate tools fluently while ignoring privacy, authorship or social consequence.
These are different profiles.
Assessment should therefore preserve dimensionality long enough to be educationally useful.
Knowledge can be assessed through explanation and conceptual tasks. Skills often require performance in realistic contexts. Attitudes may require scenario-based judgment, reflection and observation over time rather than simple agreement scales.
A total score can be created later if there is evidence that such aggregation is valid for a particular use.
Starting with a single score because dashboards prefer one number would reverse the logic of the framework.
Teacher AI literacy is an enabling condition
A student framework cannot succeed if teachers are asked to teach capacities they have not been supported to develop.
Teacher competence includes more than software familiarity.
Educators need to understand when AI output is unreliable, how to design assignments that preserve learning, how to discuss uncertainty, how to protect student data, how to identify accessibility benefits and risks, and how to respond when students use AI in ways the school has not anticipated.
Professional development also needs subject specificity.
AI use in mathematics creates different verification problems from AI use in history, language learning or visual arts.
A generic “AI day” may raise awareness without creating durable pedagogical capacity.
The framework can guide teacher learning, but it does not supply the implementation infrastructure by itself.
Assessment design becomes part of AI literacy policy
Generative AI changes what counts as evidence that a learner can perform a task.
If a student submits a polished essay, the artifact no longer reveals clearly which parts of planning, argumentation, retrieval, drafting and revision were performed independently.
Schools therefore need assessments that make process visible where process matters.
This might include oral defense, in-class components, version histories, reflective explanation, source trails or tasks that require learners to critique AI output rather than simply produce a final product.
None of these methods is universally superior.
They create trade-offs in teacher workload, accessibility and validity.
The important principle is alignment: if the learning goal is independent reasoning, the assessment must contain some evidence of independent reasoning.
The framework should be tested for transfer across systems
European education systems differ in curriculum structure, teacher autonomy, digital infrastructure, assessment culture and access to AI tools.
A common framework is valuable precisely because it creates shared language across those differences.
But shared language should not be mistaken for identical effects.
Implementation studies should examine whether the same domain has equivalent meaning across age groups, languages and national systems. Translation quality matters. So does whether students encounter commercial AI daily or only inside school-controlled environments.
Nordic systems may bring strong media-literacy traditions and high digital access, yet these conditions can create their own assumptions. Familiarity with technology is not evidence of calibrated trust.
Cross-national evaluation should look for both common competence and context-specific expression.
Equity questions begin before the lesson
AI literacy is often framed as protection against risk and preparation for opportunity.
Both depend on access.
Some learners have paid tools, private tutoring, powerful devices and parents who use AI professionally. Others encounter AI through free systems, shared devices or automated decisions they did not choose.
Accessibility can also create genuine benefit: translation, speech support, alternative explanations and assistive interfaces may expand participation for some students.
A school policy that simply bans or celebrates AI can therefore distribute costs unevenly.
The framework is strongest when it prompts institutions to ask who gains, who bears risk and whose experience is missing from implementation data.
AI literacy should increase agency, not become another marker of unequal access to expertise.
What the evidence does not show
The 2026 framework does not show that:
- students who receive AI-literacy instruction will automatically become better critical thinkers;
- the four domains constitute four validated psychological factors;
- one country, curriculum or age group has already established the best implementation model;
- using AI more frequently produces greater AI literacy;
- students should use generative AI in every subject;
- AI literacy can be adequately measured with a short knowledge quiz;
- schools can transfer responsibility for safe AI use onto individual children;
- the framework resolves current scientific disagreements about learning with AI.
Those claims require separate evidence.
Institute synthesis: AI literacy as capacity infrastructure
The Institute's synthesis is that the framework is most useful when read as capacity infrastructure rather than a checklist of software competencies.
The domains describe forms of human capability that matter because AI changes the environment in which attention, judgment, authorship and responsibility are exercised.
That interpretation should remain clearly marked as synthesis.
The OECD and European Commission framework does not use the Institute's Inner Technology terminology, nor does the Institute claim ownership of its competence structure.
The bridge is functional: both ask what people need to be able to notice, understand, judge and do in an environment where external intelligence is increasingly available.
From framework to classroom evidence
The next scientific questions are implementation questions.
Which learning activities build durable competence rather than temporary familiarity?
What transfers from one AI system to another?
How should teachers distinguish desirable assistance from cognitive offloading that weakens learning?
What forms of assessment remain valid when AI can participate in the assessment itself?
How do accessibility, language, socioeconomic inequality and teacher workload affect implementation?
And which capacities should remain deliberately practiced without AI so that students retain the ability to think and act when the system is unavailable, inappropriate or wrong?
The framework does not close those questions.
It makes them harder to avoid.
The student at the desk still has the fluent answer in front of them.
AI literacy begins before they accept it and continues after they have used it.
The important competence is not merely knowing how to make the system respond.
It is knowing what kind of human judgment is still required when it does.
Research references
- OECD / European Commission. Empowering Learners for the Age of AI: An AI Literacy Framework for Primary and Secondary Education. OECD Publishing; 2026. https://doi.org/10.1787/65cd27d4-en
- European Commission. New AI literacy framework helps schools prepare learners for the age of artificial intelligence. 2026. https://education.ec.europa.eu/whats-new/news/new-ai-literacy-framework-helps-schools-prepare-learners-for-the-age-of-artificial-intelligence
- Lee H-PH, Sarkar A, Tankelevitch L, et al. The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. CHI 2025. https://doi.org/10.1145/3706598.3713778
- Fischer H, Huff M. Metacognition as a target of science communication. Nature Reviews Psychology. 2025;4:554-555. https://doi.org/10.1038/s44159-025-00478-2
- Maertens R, Götz FM, Golino H, et al. The Misinformation Susceptibility Test (MIST): A psychometrically validated measure of news veracity discernment. Behavior Research Methods. 2024;56:1863-1899. https://doi.org/10.3758/s13428-023-02124-2
- European Commission. Guidelines for teachers and educators on tackling disinformation and promoting digital literacy through education and training. Updated 2026. https://education.ec.europa.eu/focus-topics/digital-education/actions/plan/guidelines-for-teachers-to-foster-digital-literacy-and-tackle-disinformation