What AI Literacy Actually Requires
AI literacy is not one ability hidden behind a confidence rating. It is a structure of knowledge, human skills, attitudes, and practical competences that help a learner decide what an AI system is doing and what role it should have in a task. This chapter uses the OECD/European Union AILit Framework to make that structure visible, so readers can recognise the evidence expected when someone engages with, creates with, manages, or shapes AI.
The framework matters because it prevents a useful but narrow question—“Can this person operate the tool?”—from standing in for a much larger one: “Can this person use the tool while preserving understanding, judgement, and responsibility?” A learner may know how to write an effective prompt and still lack the knowledge needed to interpret an answer, the skill needed to test it, or the disposition needed to stop using it when the task calls for human judgement.
Three layers keep literacy from collapsing into tool skill
The AILit Framework begins with three layers: Knowledge, Skills, and Attitudes. They are not three names for the same thing. They are different parts of a working system, and each changes what the others can accomplish.
Knowledge gives a learner a model of the object they are using. The framework expects learners to understand that AI systems combine procedures and statistical inferences to process data, detect patterns, and produce predictions, content, or recommendations. Generative AI can create advanced text, images, audio, and other outputs, but it does not possess authentic human understanding or intent. Its outputs may vary with the input and with design choices, and the system may be more or less transparent about how it reaches them.
This knowledge does not have to turn every learner into a machine-learning specialist. Its purpose is practical. If a learner understands that an answer is generated from patterns rather than from human comprehension, they have a reason to treat a fluent explanation as something to evaluate. If they know that data and design choices shape outputs, they have a reason to ask what perspectives may be missing or amplified. If they know that AI systems differ, they have less reason to assume that success with one tool transfers automatically to another.
Skills determine what the learner can do with that model. The framework applies critical thinking, collaboration, creativity, problem solving, computational thinking, communication, and self and social awareness to AI contexts. Critical thinking includes checking the accuracy, relevance, fairness, and bias of an output. Problem solving includes deciding whether AI is appropriate before using it, comparing approaches, monitoring progress, and changing course. Communication includes describing how a system works without presenting it as a human mind. Self and social awareness includes noticing how AI affects a learner, a community, relationships, and the environment.
Attitudes determine how the learner approaches uncertainty and consequence. The framework names reflection, responsibility, curiosity, innovation, adaptability, and empathy. These are not decorative personality traits. A reflective learner tests claims instead of accepting a persuasive narrative. A responsible learner considers unintended effects and recognises accountability for their choices. An adaptable learner can reframe a problem after an unpredictable or biased output. An empathetic learner asks who may benefit or be harmed, including people who are not present in the interaction.
The layers become meaningful when they meet in one decision. Imagine a learner asking an AI system to help prepare a presentation. Knowledge tells them that the generated material may be plausible without being a reliable account of the subject. Critical thinking gives them ways to compare claims with appropriate evidence. Communication helps them specify the purpose, audience, and constraints of the task. Self-awareness prompts them to notice whether the tool is clarifying their own thinking or replacing it. Responsibility may lead them to remove private information or decline to use generated material that cannot be properly attributed. Confidence can support the attempt, but it does not perform any of these functions by itself.
This is also why ethics is woven through the framework rather than attached as a final warning. The framework links AI literacy to learner agency, transparency and explainability, fairness, inclusion, accountability, privacy, environmental stewardship, and inquiry into who benefits or is disadvantaged. A technically capable user may still make an irresponsible decision about data or authorship. An enthusiastic user may still need to learn that choosing not to use AI can be the more informed option. Literacy is demonstrated when values affect action before, during, and after interaction.
The three layers therefore describe dependencies. Skills without knowledge can produce energetic but poorly calibrated use. Knowledge without skills can remain a correct definition that never changes a decision. Attitudes without either may express good intentions without a reliable way to act on them. A confidence question compresses these differences into a single subjective impression and cannot show where the missing layer lies.
The framework is a normative account of what learners should develop, not a claim that one questionnaire has already measured all of it. It offers a vocabulary for identifying intended outcomes and encourages educators to adapt them to local contexts. That distinction matters: a well-designed framework can tell us what evidence to seek without pretending that the evidence has already been collected.
Four domains turn qualities into decisions
The framework’s four domains show how the layers become practical. They are presented in a sequence that reflects a possible learning path, but they are better understood as connected forms of agency than as four isolated subjects.
Engage with AI: recognise, explain, and evaluate
Engage with AI is the foundation. It asks learners to identify where AI operates, understand its opportunities and risks, evaluate outputs, and decide what role AI should play in daily life. It includes explaining how systems perform tasks without common anthropomorphic claims, examining how recommendations can narrow perspectives, considering how bias may be amplified, and analysing whether use aligns with ethical principles and human values.
The framework makes evaluation concrete. A learner begins by understanding why an AI-generated output needs verification. With more development, they check accuracy and relevance against trusted sources and the expectations of the task. At a more advanced level, they assess the output and justify whether to accept, revise, or reject it. The progress is not from “low confidence” to “high confidence.” It is from awareness of a risk to an evidence-based decision.
Suppose an AI tool proposes an interpretation of a historical event. The relevant question is not whether the learner feels comfortable with the tool or whether the interpretation sounds balanced. The learner must compare it with appropriate sources, consider the assignment’s purpose, and explain what should be retained or changed. That explanation reveals knowledge, critical thinking, and responsibility in a way that a confidence rating cannot.
Create with AI: preserve the learner’s contribution
Create with AI concerns open-ended work: exploring perspectives, visualising ideas, prototyping, eliciting feedback, and refining results. The framework does not treat the generated artefact as the whole achievement. It emphasises the learner’s own ideas and asks how those ideas developed through interaction.
That focus changes what counts as evidence. A learner who generates ten images or drafts may be highly active but may not be making meaningful creative decisions. A stronger demonstration would show an original direction, compare suggestions with that direction, explain which elements were adapted or rejected, and consider authenticity, attribution, permission, and intellectual property. The final product matters, but the learner’s reasoning about the product matters too.
This domain also exposes a common confusion between confidence and creativity. A person may confidently accept an AI suggestion because it is polished, while a more tentative person may use it as a starting point and transform it substantially. Neither confidence nor the number of outputs tells us how much agency the learner retained. We need to see how they direct, interpret, and revise the contribution.
Manage AI: decide where the system belongs
Manage AI is the domain most directly concerned with allocation. Learners decide whether AI suits the task, compare possible approaches, decompose a problem, assign steps between human strengths and AI capabilities, and monitor progress against success criteria. The framework’s advanced expectations include the task’s complexity, the need for human judgement, data quality, efficiency, transparency, desired outcomes, and potential impacts.
Consider writing an essay. Choosing a topic, researching evidence, organising an argument, drafting paragraphs, revising, and proofreading do not have identical requirements. Some steps may benefit from AI assistance; others may depend on the learner’s own voice, interpretation, or accountability. A literate learner can explain that distribution rather than simply announcing that they are confident using AI for the whole assignment. They can also change the distribution when the output is inaccurate, off-target, or inconsistent with the purpose of the work.
Management is therefore not a command to use AI more efficiently. It is a judgement about where delegation helps and where it weakens learning or responsibility. The same tool may be appropriate for brainstorming in one task, inappropriate for an emotionally sensitive decision, and useful only under close checking in a research task. A learner who knows when not to use it is showing more agency, not less technological competence.
Shape AI: understand that systems can be questioned and improved
Shape AI extends the learner’s view from using a system to examining how systems are made and how they affect people. The framework asks learners to investigate a system’s purpose, intended users, and limitations; evaluate it using criteria, expected outcomes, test cases, and user feedback; consider how data selection and information flow influence behaviour; and propose improvements that support human well-being and societal benefit.
This does not require every learner to build a commercial AI product. The educational point is that systems are not inevitable forces of nature. People choose objectives, collect and organise data, define categories, evaluate performance, and decide how outputs reach users. Those choices can make a system more useful for some groups and less fair or accessible for others. Even a simple exercise comparing recommendations for different users can help a learner see that the result depends on what the system was designed to notice.
Shape AI is the culminating domain because it depends on the earlier ones. A learner cannot sensibly propose an improvement without some understanding of how the system works, some experience evaluating its behaviour, and some ability to consider consequences. Confidence may encourage experimentation, but it does not by itself provide a basis for changing a system or judging whether a change improves it.
Progression is evidence of changing ability, not a confidence ladder
The framework accompanies each competence with Basic, Intermediate, and Advanced learner expectations. These levels are not tied to a fixed age or grade, because learners begin with different experience and background knowledge. They are intended to show how the quality of understanding and action can develop.
At a basic level, a learner may identify that AI is present, recognise that an output needs checking, or identify a task for which AI might help. At an intermediate level, they can describe a specific system, compare an output with trusted information, break a problem into steps, and redirect the system when needed. At an advanced level, they can justify decisions using task criteria, weigh trade-offs, allocate roles deliberately, design evaluation criteria, or propose improvements with technical and ethical constraints in view.
These examples make a measurement principle visible. Development is not just an increase in how often learners use AI or how confident they feel while using it. It is a change in what they can explain, compare, justify, and adapt. An assessment can therefore ask a learner to inspect an output, state what they think is reliable, identify what requires checking, defend a decision, and revise the workflow after feedback. Such a task samples several parts of literacy at once without pretending that one performance captures every context.
The framework does not supply a universally validated scoring protocol. It provides competences and expectations from which educators can build context-appropriate evaluation. That boundary is important for the central question. The framework strongly supports the claim that confidence alone is too narrow a construct, but it does not provide a statistical conversion from confidence to literacy. The defensible move is to use its expectations as criteria for seeking evidence, then remain honest about what a particular task can and cannot show.
The result is a more useful mental model. AI literacy is not a feeling that sits inside a learner and can be read from a single number. It is a relationship among a learner’s understanding of the system, their human skills, their dispositions, and the decisions they make in context. The framework tells us what those decisions should involve: evaluating an output, preserving agency in creation, allocating work deliberately, and questioning how systems affect people.
Whether this model describes real behaviour is a separate question. A framework can name the capabilities that matter, but learners may enact them unevenly. They may use AI fluently for routine requests and rarely verify it; they may become highly reflective only when something goes wrong; they may judge an answer without understanding the system behind it. To see that variation, we have to move from intended competence to the traces of actual interaction—and examine what confidence hides when the system answers smoothly.