The prompt works.
A spreadsheet is analysed. A summary arrives. The language is polished enough that nobody in the meeting asks how the model reached the conclusion.
The danger in that moment is not that the tool is artificial.
It is that fluency can make competence look transferable.
A person who knows how to obtain a useful output does not automatically know what the system knows, where it fails, when verification is necessary, how bias enters, what information should not be uploaded, or who remains accountable for the decision.
That wider capacity is the territory of AI literacy.
Definition
For this Knowledge Library, AI literacy means:
the knowledge, skills and dispositions needed to understand AI systems at an appropriate level, evaluate their outputs and limitations, interact with them effectively and safely, and make responsible human judgments about when and how they should be used.
The wording is deliberately broader than technical proficiency.
AI literacy includes conceptual understanding, critical evaluation, practical interaction, ethical awareness and responsibility.
It is also a moving target because AI systems, interfaces, institutions and uses are changing quickly.
The field predates the generative-AI boom
Duri Long and Brian Magerko’s influential 2020 CHI paper defined AI literacy as a set of competencies enabling people to critically evaluate AI technologies, communicate and collaborate effectively with AI, and use AI in different settings.
Their paper synthesized interdisciplinary literature into a conceptual framework and proposed design considerations for learner-centred AI education.
It was not a validated universal measurement instrument.
That distinction matters. The paper helped give the field a shared vocabulary, but the field has continued to expand as generative systems moved into everyday work, education and public services.
AI literacy is not prompt literacy
Prompting is useful.
It is one interaction skill among many.
A person may be excellent at eliciting fluent text while knowing little about data provenance, probabilistic generation, hallucination, evaluation, privacy or automation bias.
Conversely, someone may understand the limitations of a model well enough to make wise decisions without being a sophisticated prompt engineer.
Reducing AI literacy to prompt craft creates a peculiar form of dependence: people become skilled at operating an interface while remaining unable to judge the epistemic quality of what comes back.
Technical depth should be proportionate to role
AI literacy does not require every citizen to become a machine-learning engineer.
Different roles need different levels of technical understanding.
A software developer, teacher, doctor, journalist, procurement officer and teenager do not need identical knowledge.
But role differences do not eliminate common questions:
What kind of system is this? What is it optimized to do? What are plausible failure modes? What data or context does it lack? What should be verified? What consequences follow if it is wrong? Who is accountable?
A literacy framework should therefore distinguish foundational public competence from professional and specialist competence.
Understanding capability without anthropomorphism
One central challenge is language.
People readily describe AI systems as knowing, understanding, deciding, wanting or reasoning.
Sometimes those words are useful shorthand. They can also invite false assumptions about mechanism and reliability.
AI literacy does not require banning ordinary language. It requires understanding that conversational form is not evidence of human-like cognition, memory, intention or epistemic responsibility.
A system can produce a persuasive explanation without possessing the kind of grounded understanding a reader may attribute to it.
This gap between interface impression and system capability is one of the most important literacy problems in generative AI.
Evaluation is a core competence
AI output should be evaluated in relation to task, evidence and consequence.
A brainstorming suggestion has a different verification burden from a medical recommendation, legal interpretation or public factual claim.
Evaluation can involve checking sources, reproducing calculations, comparing outputs, inspecting uncertainty, testing edge cases or consulting domain expertise.
The point is not to verify every sentence manually.
It is to calibrate scrutiny to risk.
The higher the consequence and the less independently inspectable the system, the stronger the need for human verification and institutional controls.
The confidence problem
Generative AI adds a specific metacognitive difficulty: the output often does not look uncertain when the underlying support is weak.
Human readers use linguistic fluency as a cue. In ordinary communication that can sometimes correlate with expertise. In generated text it can be misleading.
The 2025 CHI study by Lee and colleagues is relevant here. In a survey of 319 knowledge workers and 936 examples of GenAI use, higher confidence in the AI was associated with lower self-reported critical-thinking effort, while higher confidence in one’s own task ability was associated with more.
This is not causal evidence that trusting AI damages cognition.
It does show why confidence calibration belongs inside AI literacy.
European frameworks now treat AI competence as transversal
DigComp 3.0, published by the European Commission Joint Research Centre in November 2025, integrates AI competence across all 21 digital competences rather than placing it in a small isolated box.
That is an important architectural choice.
AI increasingly affects how information is searched, created, evaluated, communicated and acted upon. A person can no longer be digitally competent while treating AI as entirely separate from information literacy, media literacy, privacy or problem solving.
DigComp is a competence framework, not an intervention trial.
It describes what European digital competence should include. It does not prove that any particular curriculum successfully develops those competences.
UNESCO frameworks add human-centred and ethical dimensions
UNESCO’s 2024 AI competency frameworks for students and teachers similarly organize AI competence beyond operational use.
They emphasize human-centred understanding, ethics, AI techniques and application, and pedagogical or design dimensions appropriate to the audience.
Again, these are normative educational frameworks rather than psychometric proof of a single latent construct.
Their convergence with HCI and European digital-competence work is nevertheless informative: AI literacy is consistently being treated as a combination of understanding, evaluation, use and responsibility.
Regulation has made AI literacy an organisational concern
The EU AI Act brought AI literacy into institutional practice.
Article 4 began applying on 2 February 2025. As of September 2026, the obligation remains for providers and deployers to take measures supporting a sufficient level of AI literacy among staff and others dealing with AI systems on their behalf, while the Commission’s implementation materials emphasize context, role and risk rather than prescribing one universal certificate or score.
This is a legal and organisational requirement, not evidence that AI literacy has a settled scientific measurement model.
The distinction matters because compliance language can create false precision.
An organisation cannot infer genuine competence simply because employees completed a short module.
Measurement is still developing
Researchers have proposed AI literacy scales, competency taxonomies and scenario-based measures.
The field is young enough that construct boundaries and validation remain active questions.
Self-report measures are convenient but vulnerable to calibration problems: people may overestimate what they understand, especially in a field where terminology changes quickly.
Knowledge tests can assess concepts but may miss judgment under pressure.
Performance tasks can capture applied competence but are often tied to specific tools that age rapidly.
A serious measurement programme therefore needs multiple methods and explicit statements about what an instrument does not capture.
Literacy can become obsolete if it teaches interfaces instead of principles
Tool-specific instruction has a short half-life.
Menus change. Models improve. Product policies shift. Features disappear.
Durable AI literacy must therefore include principles that survive interface turnover:
provenance, uncertainty, task decomposition, verification, data sensitivity, accountability, bias, human oversight and the difference between generated plausibility and warranted belief.
Specific tool skills still matter. They should sit on top of those principles rather than replacing them.
AI literacy and epistemic agency
The deepest link in this Knowledge Library is not technical.
It is about who remains capable of knowing.
A person can be highly AI literate yet choose to delegate appropriately. Epistemic agency does not require doing everything manually.
It requires knowing when delegation is acceptable and retaining enough understanding to supervise the process.
AI literacy supports that agency by making the system legible enough for judgment.
Without it, delegation can become surrender without awareness.
AI literacy includes knowing when not to use AI
Competence is often demonstrated by successful use.
Sometimes it is demonstrated by refusal.
A person may decide that a task contains information too sensitive to place in a particular system, that the model cannot provide the provenance needed for a high-stakes claim, or that automation would remove a learning process the task was designed to cultivate.
These are not anti-technology decisions. They are literacy decisions.
A mature competence framework needs negative capability: the ability to recognize situations in which convenience is not the governing value.
Models, products and systems should not be conflated
Users often say “the AI” as though one stable entity produced the outcome.
In practice, an AI product can include a foundation model, retrieval system, safety layer, tools, memory features, system prompts, user settings and human review.
A model update can change behaviour without changing the product name. A product can route tasks to different models. Retrieval can make one answer current while another relies on older learned patterns.
AI literacy therefore includes a basic distinction between model capability and system behaviour.
This helps users understand why observations about one interface version do not automatically generalize to every AI system.
Calibration is more important than confidence
The goal is not to make users confident about AI.
Nor is it to make them afraid of it.
The goal is calibration.
A well-calibrated user has roughly appropriate trust for the task and system. They know which outputs are low-risk, which require checking, and which should not be delegated.
Calibration is dynamic. A system can be excellent at summarizing a provided document and poor at reconstructing an obscure citation. A user can be expert in one domain and novice in another.
This is why a universal attitude such as “trust AI” or “never trust AI” is intellectually weak.
Learning with AI versus performing with AI
Education creates a special problem.
A tool can improve task performance while reducing the amount of practice a learner performs themselves.
That does not mean assistance is inherently bad for learning. Scaffolding has always been part of education.
It means the outcome has to be named.
Did the learner produce a better essay while assisted? Did they understand the material better afterward without assistance? Did they become better at asking questions, evaluating evidence or transferring the concept to a new problem?
AI literacy for education therefore includes understanding the difference between assisted performance and retained learning.
Organisational AI literacy is distributed
An organization’s competence is not the average of individual quiz scores.
Responsible use depends on governance, procurement, documentation, security, legal interpretation, training, escalation routes and domain expertise.
A frontline employee may know a tool’s practical failure modes better than the procurement team. A data-protection specialist may understand risks that users do not. Managers may control incentives that determine whether people report errors.
This makes organizational AI literacy a distributed capacity.
Article 4 implementation should therefore be read in context: role-sensitive measures matter because the knowledge needed to use AI responsibly is not located in one person or one training module.
AI literacy and power
Literacy frameworks can become overly individualistic.
Users are told to understand systems whose design, data and business models remain opaque to them.
There is a limit to what individual competence can compensate for.
Providers and institutions retain responsibilities for transparency, safety, documentation and appropriate use.
AI literacy should empower people to ask better questions and recognize limitations. It should not become a way of transferring all risk to users after the system has been deployed.
That boundary is especially important in workplaces and public services where individuals may not have meaningful choice about whether to use a system.
Evidence of competence should be task-based where possible
If AI literacy is measured only by vocabulary, institutions may certify people who can define hallucination but cannot detect when one matters.
Task-based assessment can ask users to compare outputs, identify missing provenance, choose an appropriate verification method, recognize when sensitive data should not be submitted, or decide when human escalation is required.
These tasks still need validation. They can become obsolete when products change, and scoring may embed assumptions about “correct” use.
The principle is narrower: competence claims should include evidence of applied judgment, not merely course completion or self-reported confidence.
That principle also improves training design. When learners have to justify why they trusted, checked or rejected an output, educators can see the reasoning process rather than only the final answer.
What the evidence does not show
The evidence does not show that AI literacy is one settled psychological construct with a universally accepted measure.
It does not show that prompt skill is a valid proxy for broad AI competence.
It does not show that every user needs the same technical depth.
It does not show that the EU legal obligation can be satisfied by one standardized course or test.
It does not show that AI use necessarily reduces critical thinking.
And competence frameworks from UNESCO, DigComp or other institutions should not be presented as causal intervention evidence.
What this changes for Inner Technology
AI literacy is the interface layer between external intelligence and human discernment.
The Institute therefore treats mature AI use as requiring a recurring sequence:
Understand the system enough for the task. Specify what is being delegated. Inspect what cannot be seen directly. Verify according to consequence. Protect sensitive context. Keep uncertainty visible. Retain responsibility for the final judgment.
This sequence is Institute synthesis, not a validated AI-literacy score.
Its purpose is to make a rapidly developing field operational inside a wider architecture of human capacity.
After the fluent answer
The meeting returns to the spreadsheet.
The AI’s summary may be correct.
Someone now asks where two numbers came from. Another checks the underlying source. A third notices that the model collapsed two categories that matter operationally.
The tool has not failed.
The humans have begun using it literately.
AI literacy is not demonstrated by how quickly the machine responds.
It is demonstrated by whether people remain capable of deciding what the response is worth.
Research references
- Long D, Magerko B. What is AI Literacy? Competencies and Design Considerations. CHI 2020. https://doi.org/10.1145/3313831.3376727
- Cosgrove J, Cachia R. DigComp 3.0: European Digital Competence Framework, fifth edition. European Commission Joint Research Centre. 2025. https://doi.org/10.2760/0001149
- UNESCO. AI Competency Framework for Students. 2024. https://www.unesco.org/en/articles/ai-competency-framework-students
- UNESCO. AI Competency Framework for Teachers. 2024. https://www.unesco.org/en/articles/ai-competency-framework-teachers
- Lee HP, Sarkar A, Tankelevitch L, Drosos I, Rintel S, Banks R, Wilson N. The Impact of Generative AI on Critical Thinking. CHI 2025. https://doi.org/10.1145/3706598.3713778
- European Commission. AI talent, skills and literacy / Article 4 implementation materials. Current through September 2026. https://digital-strategy.ec.europa.eu/