Rethinking Generative AI Literacy: An Integrative, Developmental, and Dialectical Framework for K-12 Teacher Education

AI summary

Overview

Research area: AI literacy and K–12 teacher education, positioned within AI safety and ethics (arXiv category cs.CY).

Technical level: Beginner-friendly. The paper is conceptual and theory-building rather than technical; it does not report model benchmarks, datasets, or experiments. The GenAI mechanics it discusses (transformer training, reinforcement learning from human feedback, hallucination, bias) are described at an explanatory level rather than a mathematical one.

Scope in one sentence: The paper proposes RAIL-Ed (Responsible AI Literacy in Education), a conceptual framework of six interdependent pillars for preparing K–12 pre-service and in-service teachers to use generative AI responsibly.

What This Paper Is About

Generative AI tools such as ChatGPT, Claude, Gemini, and Copilot entered classrooms faster than teachers were prepared to use them well, producing what the authors call a "GenAI literacy lag": technological diffusion outpaces educators' conceptual, pedagogical, and ethical readiness. Existing AI literacy frameworks were largely built for earlier, predictive and rule-based systems, and while newer work touches on ethics, prompt literacy, and critical evaluation, it remains fragmented and rarely designed specifically for K–12 teachers. The paper's goal is to synthesize the literature into a single theoretically grounded framework that treats ethics, equity, and agency as constitutive parts of literacy rather than add-on principles.

Key Contributions

1. The RAIL-Ed framework itself. Six interdependent pillars are specified: Technical Fluency, Critical Evaluation, Human–AI Collaboration, Contextual Awareness, Ethical Reasoning, and Empowered Agency. The framework is grounded in four traditions: Freire's critical pedagogy, Dewey's reflective inquiry, Vygotsky's sociocultural theory, and Shneiderman's human-centered AI principles.

2. Three governing commitments that distinguish the framework. It is integrative (removing any pillar produces a characteristic pedagogical failure), developmental (a three-level rubric — Emerging, Competent, Advanced — describes how each pillar matures across the K–12 teacher-preparation continuum), and dialectical (the same generative affordance can deepen or displace learning depending on the literacy the teacher brings, making literacy cultivation, not tool adoption, the object of design).

3. A documented gap analysis. A systematic review and qualitative framework analysis of 67 studies (2023–2025), coded against five analytic dimensions (knowledge, skill, ethics, equity, agency) and compared against five leading frameworks, showing where the literature concentrates and where it is thin.

4. Alignment with international policy standards. Explicit mapping of RAIL-Ed's pillars onto the UNESCO AI Competency Framework for Teachers (Miao & Cukurova, 2024) and the OECD/European Commission AILit framework (2026), plus a comparison table situating RAIL-Ed among ten widely cited frameworks.

Main Findings

• The literature concentrates on knowledge and skill, not equity or agency. Across the 67 reviewed studies (2023–2025), Knowledge was addressed by 65 studies (97%), Skill by 64 (96%), Ethics by 66 (99%), Equity by 49 (73%), and Agency by 40 (60%). Equity and agency were rated as "thin" in depth of treatment; ethics was rated "present but fragmented."

• Existing frameworks predate large language models. Long and Magerko's (2020) design-centered model (17 discrete competencies) and Ng et al.'s (2021) four-domain construct (know and understand AI, use and apply AI, evaluate and create AI, AI ethics) are the most frequently cited in the corpus, and neither was designed for the demands tools such as ChatGPT, Claude, and Gemini place on K–12 teachers.

• Recent GenAI-specific work is still fragmented. Chiu et al.'s (2024) five-component framework (Technology, Impact, Ethics, Collaboration, Self-Reflection) and Allen and Kendeou's (2024) ED-AI Lit framework (knowledge, evaluation, collaboration, contextualization, autonomy, ethics) move toward integration, but the paper argues neither addresses generative-specific competencies, is designed for K–12 teachers, or treats equity as constitutive.

• Agency is narrowed to tool adoption. In the reviewed studies, agency is typically defined as behavioral intention to adopt AI tools, write effective prompts, and make individual tool-use decisions, rather than the ethical and civic demands of GenAI integration. The paper cites Zhang et al. (2025) reporting pre-service teachers' over-reliance and less critical agency when practicing with GenAI tools.

• Prompt literacy is unnamed as a competency. Wang et al. (2025) identified prompt literacy as emergent but underdeveloped; Yang and Appleget (2025) found most pre-service teacher prompts were broad and not purposefully structured. Barbieri and Nguyen (2025) found structured prompt training was associated with significantly higher self-efficacy and more purposeful AI use during practicum placements, while Şimşek (2025) documented considerable variation in prompt quality among practicing and pre-service mathematics teachers.

• Ethics is treated as compliance rather than design. Academic dishonesty was the most frequently identified misuse risk across the corpus, while hallucination, confabulation, representational bias in training data, and epistemic risk received far less systematic attention. Feldman-Maggor et al. (2025) identified hallucination and gender-racial representational bias in ChatGPT outputs as challenges for chemistry teacher education, arguing content expertise is necessary but insufficient to detect representational bias.

• Policy standards state destinations, not developmental pathways. UNESCO's AI Competency Framework for Teachers specifies 15 competencies across five aspects and three levels (Acquire, Deepen, Create); the OECD/European Commission framework organizes AI literacy into four domains (Engage, Create, Manage, Shape AI). The paper argues neither theorizes generative-specific dynamics nor supplies an account of how teachers acquire the capacity.

• The framework is conceptual and unvalidated. The authors state plainly that RAIL-Ed is not empirically validated, is not presented as a finalized product, and advances falsifiable propositions for future empirical testing.

Methodology in Plain English

The authors conducted a systematic review and qualitative framework analysis of 67 publications on AI and GenAI literacy in education published between 2023 and 2025. Each study was coded against five analytic dimensions — knowledge, skill, ethics, equity, and agency — drawn from concept-analysis methodology (Walker & Avant, 2005) and defined before, and independently of, the framework the paper proposes. The authors also coded the corpus against five leading existing frameworks to identify where those frameworks fall short for the generative, K–12 teacher-preparation context. They then built RAIL-Ed by synthesizing four theoretical traditions into a single lens: critical pedagogy (purpose), pragmatism (method), sociocultural theory (mechanism), and human-centered design (design principle). No empirical validation, survey, intervention, or experiment is reported; inter-coder agreement and detailed coding criteria are stated to appear in a companion review (Zhou et al., in press).

Why This Matters

Impact on research. The paper reframes GenAI literacy from a checklist of competencies to an interdependent, developmental, and dialectical construct, and it treats equity, ethics, and agency as constitutive rather than supplementary. It offers falsifiable propositions that other researchers can test, which the authors present as an invitation to cumulative development rather than a finished product.

Real-world applications:

• Teacher preparation programs can use the six pillars and the Emerging–Competent–Advanced rubric to structure coursework for pre-service teachers across a single developmental continuum.
• Professional development for in-service teachers can target documented weak spots, particularly prompt literacy and principled restraint, rather than general tool familiarity.
• Curriculum and assessment design can treat reflective inquiry — asking what intellectual work a tool performed, displaced, or privileged — as a structural feature of assignments rather than an end-of-task formality.
• Policy and governance work at the district or ministry level can use the mapping to UNESCO and OECD/EU standards to translate high-level commitments about human oversight, equity, and transparency into classroom-level competencies.

Industry relevance. EdTech developers, model providers, and AI governance teams working in education can read the framework as a specification of what "responsible use" demands of the human in the loop: sustained human control paired with high automation (the Shneiderman principle), explicit prompt-level intentionality, and support for principled non-use or refusal, not only adoption.

Future Directions

• Empirical validation of the framework. The authors explicitly describe RAIL-Ed as conceptual and unvalidated, and state that it advances falsifiable propositions for empirical testing — the central open task is testing whether the framework holds up in practice.

• Instrument development around the three-level rubric. The Emerging, Competent, and Advanced levels are proposed but not operationalized into measurable indicators; developing and validating assessment instruments for each pillar is an implied next step.

• Scaling beyond individual cases. The paper notes that MacDowell et al. (2024) remains an illustrative case and does not specify the conditions under which co-creative teacher engagement develops at scale across teacher preparation programs.

• Theorizing human–AI co-agency. The authors call for a theorization of co-agency grounded jointly in Shneiderman's human-centered AI principles and Freirean critical praxis, since human-centered design alone, they argue, does not address the critical consciousness teachers and students need; the OECD (2019) Learning Compass co-agency concept has not yet been systematically extended to human–AI relationships.

• Addressing under-researched risks. The paper identifies AI dependency as an under-researched moral dimension (Jorolan et al., 2025) and notes that hallucination, representational bias, and epistemic risk receive almost no pedagogical operationalization in existing frameworks.

Target Audience

This paper is most useful to teacher educators and teacher-preparation program designers, K–12 professional-development leaders, curriculum and assessment specialists, and education policy analysts working on AI integration. Education researchers studying AI literacy, critical AI literacy, and human–AI collaboration will find the gap analysis and the framework's theoretical grounding directly relevant. EdTech developers and AI governance professionals focused on education will benefit from the framework's account of what responsible classroom use requires of teachers. Because it is conceptual and requires no technical background in machine learning, it is accessible to graduate students, pre-service teachers, and administrators as well.

Authors’ abstract

Generative artificial intelligence (GenAI) has entered classrooms faster than teachers have been prepared to use it well, producing a GenAI literacy lag in which technological diffusion outpaces educators' conceptual, pedagogical, and ethical readiness. Established AI literacy frameworks predate the widespread adoption of large language models and, while acknowledging ethics, position it as a discrete competency rather than a constitutive commitment, with equity and agency as supplementary design principles. Recent GenAI-specific efforts address isolated features but remain fragmented. We introduce the Responsible AI Literacy in Education (RAIL-Ed) framework, developed through a systematic review and qualitative framework analysis of 67 studies (2023-2025), grounded in critical, pragmatist, sociocultural, and human-centered traditions (Freire, Dewey, Vygotsky, Shneiderman). RAIL-Ed specifies six interdependent pillars: Technical Fluency, Critical Evaluation, Human-AI Collaboration, Contextual Awareness, Ethical Reasoning, and Empowered Agency, marked by three commitments. It is integrative: the absence of any pillar produces a characteristic pedagogical failure. It is developmental: a three-level rubric (Emerging, Competent, Advanced) specifies how each pillar matures across the K-12 teacher-preparation continuum. It is dialectical: the same generative affordance can deepen or displace learning depending on the literacy a teacher brings to it, making the cultivation of that literacy, not the adoption of the tool, the object of design. By treating ethics, equity, and agency as constitutive, RAIL-Ed offers a theoretically grounded basis for curriculum design, teacher education, and policy, aligned with the UNESCO AI Competency Framework for Teachers and the OECD/European Commission AILit Framework. The framework is conceptual, advancing falsifiable propositions for empirical validation.