Rethinking the Structure of AI Literacy in Generative AI Educational Systems: Descriptive Patterns of Expert Convergence and Heterogeneity

Synthesis: Lee, Son, and Hong ask a question most AI Literacy frameworks skip: does expert agreement that a competency matters imply agreement about where it belongs? Working from a focused synthesis of 15 sources and a two-round Delphi study with 11 experts, they propose three domains (AI Understanding, AI Application, AI Integration) holding nine competencies, then show that judgments converge on content while dispersing on structure. Non-exclusive domain overlap (M = 4.45, SD = 0.52) and primary and secondary domain associations (M = 4.73, SD = 0.47) drew positive ratings from all 11 experts, while the proposed representation of competency placement and cross-domain applicability drew M = 3.73, SD = 1.19, IQR = 1.50. The authors read that split as grounds for treating the domains as functional rather than mutually exclusive categories. This is a conceptual and Delphi study, so no model generation applies: the generative AI the panel had in mind is described only as systems built on large language models, not as named products.

Key Findings

1. Eleven experts, all with at least 10 years of experience in SW or AI education and prior Ministry of Education or regional office policy activity, completed both rounds without attrition (100% response).
2. Round 1 competency content was rated favorably, with means from 4.27 to 4.82 and positive ratings from 81.8% to 100%; Understanding of AI Concepts and Principles scored highest (M = 4.82, SD = 0.40).
3. Structural items drew weaker convergence in Round 1: domain distinctiveness reached M = 3.91, SD = 1.22, IQR = 1.50, and cross-domain overlap M = 4.09, SD = 1.14, IQR = 2.00.
4. After between-round revision, non-exclusive domain overlap (M = 4.45, SD = 0.52) and primary and secondary domain associations (M = 4.73, SD = 0.47) received positive ratings from every expert.
5. The placement representation stayed contested at M = 3.73, SD = 1.19, IQR = 1.50, with 7 of 11 positive ratings, the lowest mean among the eight Round 2 items.
6. Experts endorsed observable behavioral indicators (M = 4.36, SD = 0.81) and human-centered agency (M = 4.27, SD = 1.19), keeping judgment, oversight, and responsibility with the human learner rather than the system.

Two Judgments, Not One

The study's analytical move is to separate two questions that competency frameworks usually fuse: is this competency important, and where does it sit? Round 1 evaluated three domains and nine competencies on a five-point validity scale, with each closed item followed by an open field. Ratings of 4 or 5 counted as positive, and consensus was deliberately not fixed to a single numerical threshold. Domain and competency ratings clustered tightly at the top of the scale. The structural items did not. The clearest contrast in Round 2 ran across three items addressing the same underlying issue: Q2 on non-exclusive overlap, Q7 on primary and secondary associations, and Q3 on the visual representation of placement. Q3 asked panelists to judge a compound figure in which nine competencies appeared as speech bubbles extending past dotted domain boundaries. It scored lowest on mean and highest on dispersion. Because it was rated collectively, the paper is explicit that the responses do not identify which individual placements generated disagreement, and it flags that differences in item scope and cognitive demand may themselves explain part of the gap.

Inside the Three Domains

The initial framework organized nine competencies under three domains, each carrying a gloss that survived into the questionnaires. AI Understanding was framed as Cognitive Extensibility and held Digital Sensitivity, Understanding of AI Concepts and Principles, and Critical Thinking. AI Application, framed as Technical Fluency, held Data Literacy, AI Tool Utilization, and Human-Centered Prompting. AI Integration, framed as Contextual Adaptability, held AI+X Convergence, Human-AI Collaboration Design, and Social Impact Evaluation. Domain means in Round 1 ran from 4.45 to 4.73, with AI Understanding highest (M = 4.73, SD = 0.65); the overarching learner concept, Human with Symbiotic Leadership, drew M = 4.55, SD = 0.93. The label work is worth noting for Theories and Frameworks: Digital Sensitivity consolidated affective and awareness constructs such as AI self-efficacy and attitudes, AI+X Convergence named the capacity to connect AI capability to disciplinary requirements, and Human-Centered Prompting was treated as an emerging competency rather than a borrowed construct.

Functional Domains and Competency Enactment

The interpretation offered is that the three domains name distinguishable functions of AI-related activity rather than containers that own competencies. AI Understanding covers making sense of systems, their capabilities, limits, and implications; AI Application covers purposeful engagement with AI on tasks; AI Integration covers connecting AI outputs to disciplinary knowledge, context, human judgment, and social considerations. The competency the authors use as a running example, Critical Thinking, has its primary association with AI Understanding, but the same judgment is required when a learner evaluates a generated response mid-task or decides whether to fold it into a disciplinary argument. From there the authors extend beyond what the Delphi measured, distinguishing competency possession from competency enactment, meaning the situated mobilization of knowledge and dispositions under task demands. Knowing that Generative AI can produce inaccurate output does not guarantee a learner will challenge a plausible but wrong response. They also propose Human-AI Collaborative AI Literacy as an interaction-oriented lens rather than a tenth competency, with agency and responsibility staying human. The Human AI Collaboration it describes is human-directed: AI capability is selectively incorporated while judgment and oversight stay with the learner. These are labeled propositions, not findings.

Curriculum, Teaching, and Assessment as Conditions

The final extension places the framework in a three-level system: competencies, activity, and educational conditions. Curriculum Design determines where opportunities for enactment appear, and the authors argue AI Literacy need not be confined to one course or unit, since AI+X Convergence only develops where disciplinary context exists inside an AI in Education program. Pedagogy and task design determine whether learners actually articulate goals and constraints, compare alternatives, and justify their use of AI output. Assessment determines which performances count, and the paper suggests performance-based tasks that show how learners evaluate outputs and allocate functions between themselves and the system, without forcing each observed behavior into one domain. Teachers mediate purposes, boundaries, and conditions of use, which connects the framework to Teacher AI Competency and to Technological Pedagogical Content Knowledge (TPACK)-style questions about integrating a capability into subject teaching rather than teaching it beside the subject. Technological conditions, including reliability and accessibility, shape what interaction is possible. The authors describe these relationships as potentially recursive rather than causal.

What this means for practice

• Instructors. Build tasks that require several competencies at once, such as asking students to state constraints before prompting, evaluate a plausible but flawed response, and justify what they kept, rather than teaching each competency as a separate objective.
• Curriculum designers. Treat AI Understanding, AI Application, and AI Integration as reference points that recur across a program, and record a primary association for each competency while leaving secondary associations visible instead of assigning competencies to exclusive units.
• Faculty developers. Prepare teachers to mediate purposes, boundaries, and verification expectations, since the framework locates the conditions for competent AI use in task design and teacher judgment rather than in tool training alone.
• Researchers. Test the placement claims the Delphi could not settle by rating each competency-domain association separately, and by collecting interaction traces or performance tasks rather than relying on expert judgment.

Limitations

• The panel was 11 experts, geographically concentrated in South Korea and skewed toward universities (n = 7, against 3 school-based members and 1 provincial office supervisor), so generalization needs more diverse panels.
• The design is descriptive rather than confirmatory: two rounds held 14 to 24 December 2025 and 5 to 15 January 2026, no consensus threshold was prespecified, and the "strong convergence" label for Q2 and Q7 is a descriptive reading of unanimous positive ratings with IQR at or below 1.00.
• Q3 rated the nine competencies collectively, so the dispersion (IQR = 1.50) cannot be attributed to specific placements, and the compound visual item may have added cognitive demand that Q2 and Q7 did not carry.
• No student-level validation was attempted; the Delphi never observed learners interacting with generative AI, so competency enactment, Human-AI Collaborative AI Literacy, and the system-level relationships remain propositions.
• The literature synthesis was purposive, resting on 7 academic studies and 8 institutional or policy frameworks, and search strings, yields, and exclusion records were not retained, so the source base should not be read as exhaustive.

Citation

Lee, S., Son, J., & Hong, J. (2026). Rethinking the Structure of AI Literacy in Generative AI Educational Systems: Descriptive Patterns of Expert Convergence and Heterogeneity. Systems, 14(10), 1231.