Demystifying AI Literacy Across Developmental Stages: An Integrative Review and the PACES Framework

Synthesis: Guo, Ali and Wong read 118 studies of AI literacy in K–12 and conclude that much of the disagreement about the term is a disagreement about the learner: a framework for a second grader and one for a senior differ in kind, not in difficulty. They sort the field into nine categories and thirty-five subcategories, trace three shifts in framework building, and turn the result into PACES, developmental considerations staged across four school bands. Their audit of validated measures finds coverage clustered in secondary schooling and nothing capturing the process of working with AI.

Key Findings

1. The competence changes in kind with the learner — early work runs on sensory play and stories; upper secondary on critique, collaboration and accountability for AI-supported decisions.
2. The field has moved from content to the learner — recent frameworks center what students can do, think and feel, folding confidence into literacy rather than treating it separately.
3. Most frameworks are written from the top down — expert syntheses dominate early childhood and primary, where classroom-tested designs are scarce; studies starting from children's ideas surface play and the human–AI relationship.
4. Ethics is usually bolted on — in knowledge-use-ethics triads and Bloom-based ladders it sits beside or after the cognitive levels, inviting a closing unit.
5. The teaching move changes with the band — stories and sensory work early, analogy and making in upper primary, simulation and critique in lower secondary, team projects later.
6. The instruments cannot see process — validated measures are knowledge tests or self-report scales, ethics is scored by agreement with principles, and none records how students question or revise AI.

Why a simpler version of the same topic is not the answer

The developmental argument is not that younger children need easier content. It is that the capacities a task loads change: a five-year-old works through symbolic play and sensory experience, a nine-year-old through concrete mechanisms, a fifteen-year-old through systems and consequences. Developmental research constrains what a task can ask, and metacognition is the clearest case, since self-assessment accuracy — which most instruments assume — is still forming in the early grades.

Two consequences follow. Children's relationship with AI is part of the curriculum: primary students treat a chatbot as a social companion before they can evaluate what it says, so epistemic vigilance has to be taught alongside the tool. And a single district-wide standard mis-serves both ends of the range.

The evidence is thinnest where the stakes are highest: five studies cover early childhood to lower primary, versus twenty-seven in upper primary.

The teaching move each band needs

In the earliest band, stories, drawing and sensory play carry the concepts, with drawing doubling as assessment because a child can show understanding before writing it.

Upper primary keeps the embodied and playful register but adds analogy and making. Students compare human and AI capacities, take roles as developer, tester and AI, and build tangible artifacts with block-based tools. Peer discussion and dilemma-based deliberation belong here too: abstract principles stay out of reach even for secondary students, while relatable dilemmas do not.

Lower secondary moves to participatory simulation: students enact an AI workflow, deliberately black-box some of its complexity, then improve a system rather than merely use it. Upper secondary organizes around agency — youth-participatory research, critique of ethics frameworks, team projects on community problems, with accountability for AI-supported decisions.

Across the bands the authors describe a rhythm rather than a ladder: build intuition early, stage critique deliberately, open technical complexity as autonomy and collaboration rise, with scaffolding shrinking but never disappearing.

Assessment sees the artifact, not the process

Nine validated measures span the bands, mostly secondary; one serves early childhood, and primary settings rely on short adapted scales. Three are knowledge tests analyzed with item response theory; the rest are self-report scales analyzed with classical test theory — a split that follows where each framework puts attitudes, not what each instrument can establish.

None captures process: how a learner steers, questions and revises with AI. Self-report dominates, with the familiar validity limits — construct under-representation, social desirability, and self-judgment that outruns metacognitive development. Ethics is measured by agreement with abstract principles, close to the opposite of judging a case.

The remedies point assessment toward performance evidence: rubrics scoring the dialogue and revision rather than the final artifact, scenario-based ethics tasks scored on reasoning as well as choice, and psychosocial items grounded in learners' actual interactions. The review also notes that curriculum aims and instruments part ways: a curriculum about critical use assessed by a multiple-choice test.

What this means for practice

• Instructors: Plan by developmental band, not by tool — story and sensory work early, analogy and making in upper primary, simulation and critique in lower secondary, team projects later.
• Program designers: Sequence AI literacy across years in the curriculum instead of adding a standalone module, and put ethics inside tasks, not in a closing unit.
• Assessment designers: Pair self-report with process evidence — dialogue rubrics, revision logs — and score ethical reasoning on the justification, not only the choice.
• Administrators: Treat PACES as a provisional roadmap; fund a primary pilot with observable indicators before adopting a district-wide sequence.

Limitations

• The search closes in late 2024, so the reviewed studies predate the current generation of generative-AI tools.
• No study follows the same learners across levels, so the progression is inferred from different cohorts rather than demonstrated.
• Evidence is uneven — five mapped studies in the first band versus twenty-seven in upper primary — and instruments are scarcer: one early-childhood tool, no fully validated primary measure.
• PACES is a design proposal: a provisional roadmap whose earliest rungs rest on the thinnest evidence and whose upper rungs assume device-rich secondary provision.

Citation

Guo, Z., Ali, M., & Wong, G. K. W. (2026). Demystifying AI Literacy Across Developmental Stages: An Integrative Review and the PACES Framework. Computers and Education: Artificial Intelligence, 100693.