Artificial intelligence literacy at school: A systematic review with a focus on psychological foundations

Synthesis: Feng and Carolus screened 2,642 records down to 58 articles on AI literacy in school education and found definitions differ by audience, with psychological elements counted far less often than knowledge elements. Their review writes separate definitions for teachers, students and other school staff and treats reasoning, self-efficacy, attitudes and emotional readiness as parts of literacy rather than preconditions. Outcomes therefore have to be written for a specific role and age, and the emotional side taught rather than assumed.

Key Findings

1. AI literacy is not one construct — teachers, students and school staff get different definitions because each group does a different job with AI.
2. The teacher version is pedagogical by construction — it joins AI knowledge to subject content and classroom constraint, so tool knowledge without a lesson does not count.
3. Two teacher dimensions describe a situation, not a score — contextual knowledge and continuous professional growth make the school's constraints and the teacher's trajectory part of the construct.
4. The student version is developmental — age appropriateness is structural, so critical evaluation means something different in primary and secondary grades.
5. Psychology sits inside literacy — reasoning, self-efficacy, attitudes and emotional readiness count as components, not preconditions learners bring or lack.
6. The literature measures what is easy to see — attitudes and reasoning are studied far more than emotional regulation, leaving the element the authors call most neglected as the thinnest evidence.

Three audiences, three definitions

Thirty-five of the fifty-eight articles offered a definition, sorted by audience — twelve for teachers, nineteen for students, four for other school roles — so the authors write three definitions rather than force them into one.

For teachers, AI literacy is the ability and self-efficacy to integrate AI knowledge with pedagogical expertise to critically analyze, apply, evaluate and develop AI technologies, aligned to subject content. Self-efficacy sits inside the definition, not beside it; thirteen reviewed articles build the account on TPACK, which supplies shared vocabulary but assumes the domains can be separated and recombined.

The student definition widens to conceptual knowledge, understanding of limits, practical skill, ethical and socio-cultural awareness, and the moral and affective dispositions that go with them. Age appropriateness becomes a dimension in its own right, alongside preparation for AI-mediated work. The definition for other roles is narrower: generic knowledge plus context-specific application, oriented to choices about tools, policy and data.

A Year 6 class and a preservice teacher therefore call for different evidence, not the same target at two levels.

What the psychological layer contains

The review sorts AI literacy into four interacting constructs — cognition, emotion, psychology and behavior — and treats behavior as emerging from the other three. The psychological dimension splits in two. The cognitive half covers reasoning and cognitive strategies, metacognition, and cognitive readiness, which adds perception, feeling, self-learning and self-assessment to knowledge preparedness. The non-cognitive half covers self-efficacy, positive attitudes, motivation and volition, and emotional readiness with its regulation and awareness components.

The counts are uneven: reasoning appears in twenty articles, attitudes in sixteen, self-efficacy and motivation in nine each, metacognition in eight, cognitive readiness in seven, emotional readiness in six. Elements a questionnaire can ask about directly get measured; regulating one's response to a confident wrong answer is the least studied.

The authors read this as support for importing self-determination theory, the theory of planned behavior and UTAUT, and recommend structured reflection, open discussion of uncertainty and ethical disagreement, and tasks that build intrinsic motivation. That is a curriculum claim, not a measurement claim: nothing shows that raising emotional readiness improves AI use, only that the field placed it in the construct.

What the dimensions ask a school to change

The teacher set adds institutional demands. Alongside AI knowledge and skills in pedagogical contexts, AI ethics and a generative-AI-specific dimension, it includes contextual knowledge and continuous professional growth — one a property of the school (devices, policy, timetabling), the other a trajectory. Together they make professional development a continuing condition of literacy rather than a certificate.

The administrator dimensions are narrower: generic knowledge plus context-specific application, oriented to decisions rather than instruction. That is a different training target, and the evidence is thin.

The authors also suggest that assessing AI literacy as a single dimension would simplify evaluation and improve cross-country comparability, which sits awkwardly beside their finding that the construct differs by role and develops with age. A composite score can compare systems; it cannot tell a school whether its teachers can place a tool in a lesson.

What this means for practice

• Instructors: Put reflection and error analysis into AI tasks — a short note on how the output felt and where it was wrong — rather than waiting for students to arrive confident.
• Program designers: Write separate outcome statements for teacher, student and administrator courses; a shared list fits none.
• Assessment designers: Measure the shared core across audiences, but score role-specific dimensions separately and treat self-report as partial evidence.
• Administrators: Fund professional development as a continuing dimension, and treat devices, policy and timetabling as parts of what teacher AI literacy has to fit.
• Researchers: Take the gap list as the agenda: empirical and longitudinal work, performance-based measures, and studies of administrators and support staff.

Limitations

• The evidence is largely conceptual, and the authors report few empirical tests of whether the dimensions matter.
• The evidence is geographically concentrated: forty-two of the fifty-eight articles came from Hong Kong and China, so the sets describe a literature more than a global practice.
• The findings on administrators, support staff and policymakers rest on five articles and are preliminary.
• The frequency counts measure how often an element appears in a framework, not how strongly it predicts anything; most reviewed instruments are self-reported.

Citation

Feng, S., & Carolus, A. (2026). Artificial intelligence literacy at school: A systematic review with a focus on psychological foundations. Computers and Education: Artificial Intelligence, 10, 100551.