Paper deep dive
Rethinking Generative AI Literacy: An Integrative, Developmental, and Dialectical Framework for K-12 Teacher Education
Shahin Hossain, Sima Ahmadi, Leqi Li, Idowu David Awoyemi, Wei Huang, Chenxi Zhou, Jujia Li, Samaa Haniya, Shapla Khanam, Tasbirun Mashreka Subaha
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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.
Tags
Links
- Source: https://arxiv.org/abs/2608.01705v1
- Canonical: https://arxiv.org/abs/2608.01705v1
Trouble viewing inline? Open PDF directly →
Full Text
111,617 characters extracted from source content.
Expand or collapse full text
Rethinking Generative AI Literacy: An Integrative, Developmental, and Dialectical Framework for K–12 Teacher Education Shahin Hossain 1 , Sima Ahmadi 2 , Leqi Li 3 , Idowu David Awoyemi 4 , Wei Huang 5 , Chenxi Zhou 6 , Jujia Li 7 , Samaa Haniya 8 , Shapla Khanam 9 , Tasbirun Mashreka Subaha 10 1 School of Education, University of Maryland Baltimore County, Baltimore, Maryland, USA • shahinh1@umbc.edu 2 School of Teaching, Learning, and Curriculum Studies: Educational Technology Department, Kent State University, Ohio, USA • sahmadi@kent.edu 3 Department of Learning and Performance Systems, Pennsylvania State University, State College, PA, USA • lzl5599@psu.edu 4 Department of Educational Leadership, Policy, and Technology Studies, The University of Alabama, Tuscaloosa, AL, USA • idawoyemi@crimson.ua.edu 5 College of Education, University of Alabama, Alabama, USA • whuang20@crimson.ua.edu 6 Graduate School of Education and Human Development, Department of Curriculum and Instruction, George Washington University, District of Columbia, USA • chenxi.zhou1@gwu.edu 7 College of Education, University of Alabama, Alabama, USA • jli183@crimson.ua.edu 8 Graduate School of Education and Psychology, Pepperdine University, Los Angeles, CA, United States • samaa.haniya@pepperdine.edu 9 Department of Artificial Intelligence, Faculty of Computer Science and Information Technology, Universiti Malaya, Kuala Lumpur 50603, Malaysia • shapla.k@um.edu.my 10 Department of English Language and Literature, Begum Rokeya University, Rangpur, Bangladesh • tasbirunmashreka.12102020@student.brur.ac.bd Abstract Generative artificial intelligence (GenAI) has entered class- rooms faster than teachers have been prepared to use it well, producing a GenAI literacy lag in which technological diffu- sion 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 treated as supplementary design principles. More recent GenAI-specific efforts address isolated features but remain fragmented. This paper introduces the Responsible AI Liter- acy in Education (RAIL-Ed) framework, developed through a systematic review and qualitative framework analysis of 67 studies (2023–2025) coded against five leading frameworks and grounded in critical, pragmatist, sociocultural, and human- centered traditions (Freire, Dewey, Vygotsky, and Shneider- man). RAIL-Ed specifies six interdependent pillars: Techni- cal Fluency, Critical Evaluation, Human–AI Collaboration, Contextual Awareness, Ethical Reasoning, and Empowered Agency, distinguished by three commitments. It is integrative: the absence of any pillar produces a characteristic pedagogi- cal failure. It is developmental: a three-level rubric (Emerg- ing, Competent, Advanced) specifies how each pillar matures across the K–12 teacher-preparation continuum. And it is dialectical: the same generative affordance can deepen or dis- place 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 Frame- work for Teachers and the OECD/European Commission AILit Framework. The framework is conceptual, advancing falsifi- able propositions for empirical validation. Keywords: Generative AI literacy; teacher education; critical AI literacy; human–AI collaboration; AI ethics in education; teacher professional development 1 Introduction The release of ChatGPT in late 2022 and the rapid prolifera- tion of generative artificial intelligence (GenAI) systems that followed mark one of the most consequential shifts in educa- tional technology in recent history (Kasneci et al., 2023; Miao & Holmes, 2023). Unlike earlier technologies such as Web 2.0, the Internet of Things, learning management systems, or rule-based intelligent tutors, generative systems such as ChatGPT, Claude, Gemini, and Copilot differ qualitatively in three ways: they generate novel content rather than retrieve it, interact through open-ended natural language rather than pre- determined rules, and produce fluent outputs whose accuracy and provenance resist verification. They generate content in real time, such as essays, code, images, audio, problem solu- tions, and arguments that learners and teachers increasingly treat as collaborative outputs (Mollick & Mollick, 2023). This shift unsettles long-standing educational assumptions about authorship, expertise, and cognition, and invites consideration of how generative systems function as epistemic mediators in learning contexts (Bender et al., 2021; Kay et al., 2024). Educational institutions therefore confront a layered respon- sibility. They must prepare learners to engage productively with generative systems while simultaneously cultivating the critical, ethical, and reflective sensibilities required to inter- rogate them (Selwyn, 2022; Miao & Holmes, 2023). Interna- tional policy standards have begun to articulate this responsi- bility through frameworks that emphasize human oversight, eq- 1 arXiv:2608.01705v1 [cs.CY] 3 Aug 2026 uity, and transparency (OECD, 2024; OECD/European Com- mission, 2026; Miao & Holmes, 2023). Translating these high-level commitments into coherent classroom practice re- mains unresolved, in part because GenAI tools evolve faster than curricula, teacher preparation programs, and institutional policy (Kasneci et al., 2023; Miao & Holmes, 2023). A growing body of scholarship documents what may be characterized as a GenAI literacy lag: the widening gap be- tween the speed at which GenAI tools enter classrooms and the slower development of the conceptual, pedagogical, and ethi- cal capabilities educators and learners need to use them well (Cheah et al., 2025; Sattelmaier & Pawlowski, 2025; Sperling et al., 2024). The result is that many users can operate these systems without being equipped to evaluate their outputs, rec- ognize their limitations, or apply them in pedagogically sound ways. Existing AI literacy frameworks, although foundational, were developed largely in response to earlier paradigms of ma- chine learning and rule-based classification (Long & Magerko, 2020; Ng et al., 2021). They tend to emphasize technical com- prehension and instrumental skill over pedagogical principle, operationalizing literacy through relatively stable, classifier- style applications that presuppose standardized baselines and competency checklists. Long and Magerko's (2020) framework, for instance, speci- fies seventeen discrete competencies, while Ng et al. (2021) suggested four dimensions of AI literacy (know and under- stand AI, use and apply AI, evaluate and create AI, and AI ethics) for students. Such models emerged largely in response to predictive, classifier-based systems, whose bounded out- puts let learners form stable mental models of how a system behaves. Generative systems introduce different demands: they depend heavily on prompt construction and rhetorical framing (Mollick & Mollick, 2023); they produce probabilis- tic and frequently inaccurate outputs commonly described as hallucinations (Bender et al., 2021; Kasneci et al., 2023); and they encode representational biases at scale (Lee et al., 2024). They can also invite forms of cognitive offloading that erode learning when used without scaffolding (Kasneci et al., 2023; Selwyn, 2022). Recent frameworks increasingly acknowledge these ten- sions individually, with growing attention to ethical reason- ing, prompt literacy, and critical evaluation of outputs (Jin et al., 2025; Ng et al., 2025; Ofosu-Asare, 2025). Although epistemic harms have been theorized in unified terms outside education (Kay et al., 2024), research rarely integrates halluci- nation, prompt dependency, epistemic risk, and AI-mediated inequity into a single theoretical framework for teaching and learning (Park, 2025). The absence of an integrative frame- work that connects technical fluency with critical, ethical, and equity-oriented capacities leaves teachers underprepared for the responsibilities GenAI imposes on practice (Lee et al., 2024; Miao & Holmes, 2023). This paper introduces the Responsible AI Literacy in Educa- tion (RAIL-Ed) framework to address this gap. We synthesized existing scholarship through a systematic review and qualita- tive framework analysis of the AI literacy literature. The frame- work's epistemological foundation draws on four traditions: three educational traditions and one human-centered design tradition. The educational traditions are Freire's (1970/2000) pedagogy of critical consciousness, Dewey's (1938) reflec- tive inquiry, and Vygotsky's (1978) sociocultural theory of mediated cognition. The fourth, Shneiderman's (2020) human- centered AI principles, which pair strong human control and oversight with high levels of automation, positions the frame- work within responsible system design. From this synthesis, RAIL-Ed articulates six interdependent pillars: Technical Flu- ency, Critical Evaluation, Human–AI Collaboration, Contex- tual Awareness, Ethical Reasoning, and Empowered Agency. What distinguishes RAIL-Ed is not its pillars alone but three commitments that govern how they operate together: it is integrative, developmental, and dialectical. This framework is intended for K–12 teacher education and is adaptable to diverse disciplinary, cultural, and institutional contexts (Ng et al., 2021; Miao & Holmes, 2023). It addresses the entire K–12 teacher-preparation continuum, encompassing both pre- service teachers in undergraduate programs and in-service teachers currently in classrooms, where the literature iden- tifies the most significant gap in preparation. These groups are treated as developmentally connected points along a sin- gle continuum. Teachers are positioned at the center of this framework because teachers’ GenAI literacy influences the literacy of every cohort of students they subsequently instruct, making teacher preparation the most impactful site for inter- vention. The scope of this work is conceptual; it does not provide empirical validation of the framework or present it as a finalized product. Instead, it offers a theoretically grounded scaffold intended to inform curriculum design, teacher educa- tion, professional development, and policy (OECD, 2024; Tan, 2025). This paper makes three primary contributions. First, it argues that responsible engagement with GenAI requires ped- agogically and theoretically grounded foundations, framing substantive learning as a deliberate design commitment. Sec- ond, it treats ethics, equity, and agency as constitutive elements of literacy, responding to concerns about representational harm (Bender et al., 2021; Lee et al., 2024), epistemic injustice (Kay et al., 2024), and structural inequity in AI-mediated learning (Selwyn, 2022). Third, it offers a framework whose dimen- sions are interdependent and mutually constitutive, conceptu- alizing literacy as a relational, situated, and ethical practice. Through these contributions, the paper advances scholarly and policy discussions on what responsible GenAI integration meaningfully requires of teachers, learners, and the broader educational community (OECD, 2024; Tan, 2025). The remainder of the paper proceeds as follows. Section 2 establishes RAIL-Ed's theoretical foundations and reviews ex- isting AI and GenAI literacy frameworks; Section 3 identifies the gaps that motivate a new framework; Section 4 presents RAIL-Ed, detailing its six pillars, failure-mode map, and devel- opmental rubric; Section 5 develops implications for practice, policy, and research; and Sections 6 and 7 address limitations, future directions, and conclusions. 2 2Theoretical Foundations: Critical, Pragma- tist, Sociocultural, and Human-Centered Tra- ditions A framework for GenAI literacy requires more than an inven- tory of competencies; it requires a theory of how teachers and learners come to know, judge, and act with a technology that mediates knowledge itself. RAIL-Ed draws on four comple- mentary traditions to supply that theory. Freire’s critical ped- agogy establishes literacy as an exercise of agency. Dewey’s pragmatism establishes it as a reflective inquiry. Vygotsky’s sociocultural psychology establishes cognition as mediated by cultural tools; and Shneiderman’s human-centered design establishes the conditions under which automation augments rather than displaces human judgment. These traditions reject the deficit thinking that has historically framed learners’ diffi- culties as personal deficiencies, and they position equity not as a topic to be added but as a property constitutive of literacy itself. The subsections that follow develop each tradition and link it to the pillars it grounds. 2.1A Critical Foundation: Rejecting Deficit Thinking (Freire) RAIL-Ed begins from a critical commitment: a rejection of the deficit thinking that attributes students’ academic difficulties to perceived deficiencies in learners, families, or communities (Davis & Museus, 2019). Such framing, prominent across twentieth-century educational discourse, justified lowered ex- pectations and remedial tracking for students from marginal- ized backgrounds (Tewell, 2020). It persists today whenever learners’ struggles with new technologies are read as indi- vidual readiness gaps. Freire’s (1970/2000) critique of the “banking model” in which teachers deposit information into passive students names the pedagogy that deficit thinking pro- duces. In opposition to it, Freire proposed a problem-posing education grounded in dialogue, reflection, and praxis, in which learners are active participants capable of transforming their conditions through critical inquiry. This commitment is directly consequential for GenAI literacy. A deficit framing re- duces literacy to operational compliance, prompt construction, and rule-following while obscuring how AI systems shape knowledge, authority, and power (Selwyn, 2022). Freire’s tradition instead requires that teachers be prepared to help students interrogate those systems, not merely operate them. It grounds three RAIL-Ed pillars: Contextual Awareness, Ethical Reasoning, and Empowered Agency, positioning equity and ethics as foundational to AI literacy. 2.2A Pragmatist Foundation:Reflective Inquiry (Dewey) If the critical tradition establishes why literacy must attend to power, the pragmatist tradition establishes how it is learned: through experience, experimentation, and reflection. Dewey (1938) understood learning as a social and democratic process in which knowledge is constructed through active inquiry and the disciplined examination of one’s own experience. Applied to GenAI, Dewey’s stance reframes the central pedagogical question from whether students can use a tool to what intel- lectual work the tool performed, displaced, or privileged in a given encounter, and what that means for the learner’s devel- oping judgment. Reflection, in this view, is not an end-of-task formality but a structural feature of practice. Grounded in crit- ical and pragmatist traditions, RAIL-Ed's Critical Evaluation pillar combines Freire's emphasis on questioning AI authority with Dewey's emphasis on reflective inquiry. 2.3A Sociocultural Foundation: Mediated Cognition (Vygotsky) While the critical and pragmatist traditions concern the pur- poses and processes of learning, Vygotsky’s (1978) sociocul- tural psychology concerns its mechanisms. For Vygotsky, higher mental functions develop through the mediation of cultural tools, language, symbol systems, and technologies that not only assist thought but also reshape it. Cognition is therefore never wholly individual; it is distributed across the learner, more capable others, and the tools through which knowledge is constructed. Generative systems represent a mediating tool of unusual power: unlike a calculator or a search engine, a large language model (LLM) participates in the formation of arguments, interpretations, and explanations, and recent scholarship has begun to theorize it explicitly as a mediational agent within sociocultural learning (Tate et al., 2026). This grounding carries a sharp implication that the framework operationalizes: a teacher who does not understand how the tool mediates the production, bias, and bounds of its outputs is mediated by it. Vygotsky's tradition thus grounds RAIL-Ed's Technical Fluency pillar, which treats functional understanding of the model as a prerequisite for control, and its Human–AI Collaboration pillar, which positions the system as a mediational agent in teaching, participating in the work without occupying the social position of a partner (Tate et al., 2026). It also informs Contextual Awareness, since what a tool mediates depends on the cultural and linguistic setting in which it is used. 2.4A Human-Centered Design Foundation (Shneider- man) The three educational traditions explain how literacy is pro- posed, learned, and mediated; a fourth specifies the design principle that should govern the human–AI relationship itself. Shneiderman’s (2020) human-centered AI rejects the common assumption that human control and machine automation trade off on a single continuum. He argues instead that they are independent dimensions, and that the most reliable, safe, and trustworthy systems occupy the quadrant of high automation and high human control at once. This reframing matters for education because it dissolves the false choice between ban- ning GenAI and surrendering pedagogical judgment to it: the goal is neither minimal automation nor minimal oversight but high levels of both, with the teacher retaining control over consequential decisions. The principle aligns with the human- in-the-loop tradition in AI-in-education research, in which human judgment retains oversight of AI-mediated processes (Memarian & Doleck, 2024), while extending it from sys- 3 tem design to pedagogical practice. Shneiderman’s tradition grounds RAIL-Ed’s Human–AI Collaboration pillar, which asks teachers to maximize the system’s generativity without ceding authority over it, and its Empowered Agency pillar, in which sustained human control is exercised through prin- cipled adoption, restraint, or refusal. Shneiderman himself resists describing AI as a teammate or partner, holding that computers should support human activity. RAIL-Ed combines Vygotskian mediation with Shneiderman's control–automation principle to support collaboration while maintaining pedagog- ical authority. 2.5 Synthesis: From Four Traditions to One Lens These four traditions are not invoked as parallel citations but as a single, integrated lens. The critical tradition supplies the framework’s purpose (literacy as agency and equity), the pragmatist tradition its method (reflective inquiry), the socio- cultural tradition its mechanism (mediated cognition), and the human-centered tradition its design principle (control and automation maximized together). Their synthesis yields the framework’s central and distinctive claim. Because cognition is mediated by tools (Vygotsky) but those tools operate within structures of power and possibility (Freire), the same gener- ative affordance does not carry a fixed value: it can deepen inquiry or displace it, widen access or entrench advantage, depending on the literacy the teacher brings to the encounter. RAIL-Ed names this its dialectical commitment, and it fol- lows from the theory. Equity cannot be treated as a separate competency because literate AI practice requires attention to the structural conditions shaping how tools are accessed and used. The pillars developed in Section 4 operationalize this lens; the gaps that necessitate such a framework are documented first in Section 3. 2.6Existing AI Literacy and GenAI Literacy Frame- works Against these foundations, existing AI and GenAI literacy frameworks can be read for what they offer and where they stop short. Long and Magerko (2020) broadened AI literacy beyond technical knowledge, defining it as a set of compe- tencies enabling individuals to critically evaluate AI, commu- nicate and collaborate with it, and use it as a tool across set- tings; their work established the human-centered turn on which later frameworks build. Allen and Kendeou’s (2024) ED-AI Lit framework advanced an interdisciplinary, six-dimensional model of knowledge, evaluation, collaboration, contextualiza- tion, autonomy, and ethics, strengthening attention to indepen- dent judgment and contextual implications. At the policy level, UNESCO’s guidance (Miao et al., 2021) foregrounded inclu- sion, equity, transparency, and human-centered development, insisting that AI must not exacerbate existing inequalities. These contributions are foundational, but they share limi- tations for the generative, teacher-preparation context. Most predate the widespread adoption of LLMs and therefore em- phasize foundational AI concepts over the distinctive demands of generative systems: prompt construction, hallucination de- tection, output verification, and human–AI co-creation (Park, 2025). They tend to position educators and learners as users rather than co-creators of AI-mediated content (Yim & Su, 2025), and they leave the ethics of generative use, academic integrity, misinformation, copyright, privacy, algorithmic bias, and data governance comparatively underdeveloped (Gruen- hagen et al., 2024). A teacher using GenAI to design as- sessments, for instance, must judge whether generated items reproduce copyrighted material, encode bias, or assert plau- sible but false content that undermines validity—judgments these frameworks do not equip them to make. Recent work in- creasingly characterizes GenAI literacy as a multidimensional competency spanning understanding, effective use, critical evaluation, ethical reasoning, and responsible creation (Park, 2025), but these conceptualizations have not yet been consol- idated into a theoretically grounded, teacher-focused frame- work. Section 3 specifies these gaps in detail, and Table 2 situates RAIL-Ed among the full set of frameworks reviewed. 3Identified Gaps and the Rationale for a New Framework 3.1 Coverage of the Reviewed Literature To establish what a new framework must provide, we ex- amined how the existing literature allocates its attention and where it falls short in the generative K–12 teacher-preparation context. We reviewed 67 publications on AI and GenAI lit- eracy in education (2023–2025); research addressing GenAI literacy specifically remains concentrated in higher education, with K–12 teacher preparation comparatively underserved. Across these studies, Long and Magerko’s (2020) design- centered model and Ng et al.’s (2021) four-domain construct are the most frequently cited frameworks, both developed be- fore large language models were publicly released at scale, and neither designed for the epistemic and pedagogical de- mands that tools such as ChatGPT, Claude, and Gemini place on K–12 teachers. To characterize where the literature concentrates, we coded each study against the five analytic dimensions that struc- tured our review: knowledge, skill, ethics, equity, and agency (Walker & Avant, 2005). Because these dimensions were de- fined a priori from concept-analysis methodology, the distribu- tion in Table 1 reports what the corpus contains independently of the framework it motivates. Coverage focuses on knowl- edge and skills, while equity and agency receive substantive treatment in a minority of studies, the systematic thinness that RAIL-Ed is designed to address. The gaps below reflect this distribution. Running through them is a single problem: existing frameworks, and even the most recent international standards, specify what teachers and learners should be able to do, but they do not provide a generative-specific, theoretically grounded, and developmental account of how K–12 teachers acquire that capacity. They name a destination without a theory of how to reach it. To make explicit how RAIL-Ed complements these stan- dards, Table 3 maps its six pillars onto the UNESCO AI Com- petency Framework for Teachers (Miao & Cukurova, 2024) 4 Figure 1: Theoretical Foundations of RAIL-Ed Framework Table 1: Coverage of Five Analytic Dimensions Across the 67 Reviewed Studies (2023–2025) Analytic dimensionWhat it capturesStudies addressing it, n (%) Typical depth of treat- ment KnowledgeHow AI and GenAI systems work; conceptual understanding 65 (97%)Foundational SkillOperating, prompting, and applying tools64 (96%)Substantive EthicsIntegrity, bias, privacy, misinformation66 (99%)Present but fragmented EquityAccess, language, structural inequality49 (73%)Thin AgencyCo-agency, principled use, restraint, civic voice40 (60%)Thin/ instrumental Note. Dimensions follow the concept-analysis matrix used in coding (Walker & Avant, 2005) and were defined prior to and independently of RAIL-Ed’s pillars. Counts reflect studies giving each dimension substantive treatment; coding criteria and inter-coder agreement are reported in the companion review (Zhou et al., in press). and the OECD/EU AI literacy framework (OECD/European Commission, 2026). RAIL-Ed operationalizes their shared commitments for the generative, K–12 teacher-preparation context, adding the developmental progression and dialectical account those frameworks leave underspecified. 3.2 Generative-Specificity As Annapureddy et al. (2025) indicated in their GenAI compe- tency model, which includes twelve items, existing AI literacy frameworks remain too generic and fail to differentiate be- tween predictive and GenAI systems. This distinction shapes what GenAI literacy means and what it requires teachers to be able to do. In another study, Sattelmaier and Pawlowski (2025) observe that K–12 teachers are navigating rapid AI- driven changes in their classrooms without a clear account of which competencies they actually need to develop. Building on the previous Gen AI literacy models, recent scholarship has proposed frameworks that aim to integrate ethics, collaboration, and interdisciplinary perspectives more centrally into AI literacy. For example, Chiu et al. (2024) developed a five-component AI literacy framework (Technol- ogy, Impact, Ethics, Collaboration, Self-Reflection) that places ethics and collaboration at the center of the GenAI framework. Similarly, Allen and Kendeou (2024) developed an ED-AI Lit framework as a holistic, interdisciplinary model of AI literacy in education. While these frameworks represent mean- ingful advances in integrating interdisciplinary and ethical dimensions of AI literacy in different subject domains, their contributions in GenAI literacy remain constrained. Neither 5 Table 2: Comparison of Widely Cited AI and GenAI Literacy Frameworks Framework (year) Focus/definitionDimensionsContextAudienceLimitation for GenAI Long & Magerko (2020) Human-centered AI literacy competen- cies 17 competencies (recog- nize AI, critical evalua- tion, communication & collaboration, using AI as a tool) General public General learn- ers Pre-LLM; prompting, hallucination, and output verification not addressed Ng et al. (2021)Conceptual review of AI literacy Four domains: know & understand, use & apply, evaluate & create, ethics General educa- tion LearnersPredictive-AI framing; positions learners as users, not co-creators AI4K12 / Touret- zky et al. (2019) K–12 AI content standards Five Big Ideas (percep- tion, representation & reasoning, learning, nat- ural interaction, societal impact) K–12 curricu- lum K–12 students Content standards, not teacher GenAI literacy; no generative focus UNESCO / Miao et al. (2021) Global AI-in- education policy guidance Inclusion, equity, trans- parency, human-centered AI International policy Policymakers Policy-level, not a classroom competency model ED-AI Lit (Allen & Kendeou, 2024) Interdisciplinary AI literacy in education Knowledge, evaluation, collaboration, contextual- ization, autonomy, ethics Education (broad) Educators & learners Not generative- specific; equity not constitutive; not K–12 teacher-focused Chiu et al. (2024)AI literacy with ethics & collabora- tion central Technology, Impact, Ethics, Collaboration, Self-Reflection School educa- tion StudentsLimited GenAI- specific competencies; no developmental ac- count Su & Yang (2023) — IDEE GenAI literacy / human-in-the-loop design Identify outcomes; De- termine automation level; Ensure ethics; Evaluate effectiveness General teach- ing EducatorsDoes not specify how teachers develop the judgment it requires Annapureddy et al. (2025) GenAI competency model 12 competencies differ- entiating generative from predictive AI Workforce/ general General usersGeneric; not K–12; no pedagogy or equity Sattelmaier & Pawlowski (2025) AI competencies for K–12 teachers Teacher-facing AI compe- tency areas K–12 schoolsK–12 teachersLimited equity / civic- agency treatment; emerging validation UNESCO AI CFT (Miao & Cukurova, 2024) Global teacher AI competency standard 5 aspects (human-centred mindset; ethics of AI; AI foundations & applica- tions; AI pedagogy; AI for professional develop- ment) × 3 levels (Acquire, Deepen, Create) K–12 teachers (policy) K–12 teachersCompetency stan- dard; not a generative- specific or developmen- tal pedagogical model UNESCO AI CF for Students (Miao et al., 2024) Global student AI competency standard 4 aspects × 3 levels (Un- derstand, Apply, Create); 12 competencies K–12 curricu- lum K–12 studentsStudent-facing; not teacher preparation OECD/EU AILit (2026) AI literacy for pri- mary & secondary education 4 domains: Engage with AI, Create with AI, Man- age AI, Shape AI (knowl- edge, skills, attitudes) K–12 school- ing Learners (teacher- mediated) Learner-outcome framework; limited teacher developmental progression AI Literacy (Mills et al., 2024) Practitioner frame- work (Digital Promise) Understand, evaluate, and use emerging AI K–12 educa- tion Educators & learners Brief, practitioner- oriented; not generative-pedagogy- specific RAIL-Ed (this paper) Responsible GenAI literacy for K–12 teacher education Six pillars: Technical Flu- ency, Critical Evaluation, Human–AI Collaboration, Contextual Awareness, Ethical Reasoning, Em- powered Agency K–12 teacher prep (pre- & in-service) K–12 teachersConceptual; awaiting empirical validation (stated openly) Note. The table characterizes each framework’s focus, dimensions, and principal limitation for GenAI; dimensions reflect each source’s own terminology. The final row presents the framework proposed in this paper. 6 model addresses the generative-specific competencies that dis- tinguish GenAI literacy from broader AI literacy, neither is designed for K–12 teachers, and neither treats equity as a constitutive design principle. Cheah et al. (2025) further dis- cussed that the lack of clear policy and instructional guidance for integrating GenAI in K–12 makes even motivated teachers ineffective in meeting practical classroom needs in real time. 3.3 Agency Narrowed to Tool Adoption An important gap in the AI and GenAI literacy theoretical frameworks of reviewed studies is the treatment of teacher and learner agency in the case of GenAI systems. The concepts of agency across the reviewed studies were consistently inter- preted in narrow, instrumental terms (Zhou et al., in press). Agency was defined as the behavioral intention of an educator to adopt AI tools (e.g., Bower et al., 2024; Kong et al., 2024), write effective prompts (Yang & Appleget, 2025), and make individual decisions about tool use (Cheah et al., 2025). This definition is practical. However, it does not fully capture the ethical and civic demands that scholars have argued must ac- company GenAI integration in education (Annapureddy et al., 2025; Laine et al., 2025; Roe et al., 2025; Tagare et al., 2025). The current AI literacy literature, and GenAI literacy studies specifically, has not yet developed a theoretically grounded account of human–AI co-agency. Although the OECD (2019) Learning Compass uses co-agency to describe collaborative relationships among learners, teachers, peers, families, and communities in support of student development, the concept has not been systematically extended to human–AI relation- ships in the educational literature. Studies that engage with human–AI collaboration tend to use the related vocabulary of co-creation (MacDowell et al., 2024; Yang & Appleget, 2025) or distributed pedagogical roles (Yu & Pian, 2025), but do not articulate co-agency as a construct grounded in a theory of learner and teacher participation in AI-mediated knowledge production. One negative consequence of this gap is blind reliance on GenAI tools. Zhang et al. (2025) reported that pre-service teachers had an over-reliance and less critical agency when they were practicing with GenAI tools. Joseph (2023) provides a concrete illustration of why this pattern is consequential. In her test of ChatGPT on a literary analysis task, the model fab- ricated factual references that students would otherwise have accepted as accurate (Joseph, 2023). Building on these empir- ical findings, Jorolan et al. (2025) argued that AI dependency is a crucial moral dimension that remains under-researched. Beyond these empirical observations, existing frameworks have largely overlooked the problem of blind reliance, treating it neither as a competency gap to address nor as an ethical risk to design against. 3.4Ethics and Equity Treated as Compliance, Not Design At the framework level, Su and Yang (2023) propose that edu- cators should determine the appropriate level of automation when integrating GenAI into teaching, a principle that aligns conceptually with the broader human-in-the-loop (HITL) tradi- tion in AI-in-education research (Memarian & Doleck, 2024), in which human judgment retains oversight over AI-mediated processes and outputs. However, their GenAI literacy IDEE framework (Identify the desired outcomes, Determine the ap- propriate level of automation, Ensure ethical considerations, Evaluate the effectiveness, Su & Yang, 2023) does not specify the conditions under which teachers develop the metacognitive capacity to exercise this judgment consistently and critically. MacDowell et al. (2024) position teachers as active creators who must develop the knowledge, skills, and mindsets neces- sary to teach with and create using GenAI in pedagogically responsible ways. This conception moves beyond tool adop- tion toward sustained engagement and co-creation. However, it remains an illustrative case, and it does not yet specify the conditions under which such co-creative engagement develops at scale across teacher preparation programs. We argue that the field needs a theorization of co-agency that draws on Shneiderman's (2020) human-centered AI prin- ciples and Freirean critical praxis (Freire, 1970/2000) alike. Such a theorization would position teachers as neither passive consumers of AI outputs nor uncritical opponents of AI tools, but instead as deliberate and ethically-engaged co-participants in AI-mediated knowledge construction. We see this dual grounding as necessary because human-centered design alone does not address the critical consciousness that teachers and students need to engage with AI on equitable terms. The other problem is the risk of misuse of GenAI tools. This is essential to how teachers and students use GenAI in classrooms. The most frequently identified misuse risk across the corpus is academic dishonesty (Bae et al., 2024; Bower et al., 2024; Bukar et al., 2024), yet this concern, legitimate as it is, functions in most studies as a compliance problem rather than as a literacy challenge. The deeper misuse risks of GenAI, such as hallucination and confabulation (Bae et al., 2024; Choi, 2025), representational bias embedded in training data (Feldman-Maggor et al., 2025), and the epistemic risk of AI-generated content that mimics authoritative knowledge while being factually unreliable, receive far less systematic attention and almost no pedagogical operationalization in the included frameworks. Feldman-Maggor et al. (2025) identi- fied hallucination and gender-racial representational bias in ChatGPT outputs as challenges for chemistry teacher educa- tion, showing that disciplinary content expertise is necessary for teachers to detect content errors and fabricated references, yet insufficient for recognizing representational bias, which demands AI-specific competencies beyond existing teacher- knowledge frameworks. This finding has broad implications for K–12 teacher preparation across subjects. 3.5 Prompt Literacy Left Unnamed Prompt literacy, which is an essential competency to construct purposeful, critically reflective, and pedagogically intentional inputs for LLMs, represents another concerning issue. Wang et al. (2025) identified prompt literacy as an emergent but underdeveloped competency that existing frameworks do not meaningfully address. A reasonable response might be that prompt literacy could be located within the knowledge or skills components of existing AI literacy frameworks: Ng et al.'s 7 (2021) "use and apply" domain, Long and Magerko's (2020) design-centered competencies, or Allen and Kendeou's (2024) Knowledge component could in principle accommodate it. In practice, however, none of these frameworks explicitly names prompt construction as a competency, operationalizes how teachers should develop it, or distinguishes it from general AI tool use. A general "knowledge of AI" or "use of AI" component is different from a framework that specifies prompt literacy as a constitutive competency of GenAI literacy. Prompt literacy is the operational bridge between teacher in- tention and AI output, and the field has documented that teach- ers without explicit prompt literacy instruction tend to default to broad, unstructured prompts that constrain the quality of what AI returns. Yang and Appleget (2025), for instance, iden- tified three levels of pre-service-teacher prompting and found that most prompts were broad and not purposefully structured. Arrington et al. (2025) proposed a four-component prompt en- gineering framework for K–12 teachers. Barbieri and Nguyen (2025) found that pre-service teachers who received structured prompt training demonstrated significantly higher self-efficacy and more purposeful AI use during practicum placements. Yet ̧Sim ̧sek (2025) documented considerable variation in the quality and intentionality of teacher-generated prompts among practicing and pre-service mathematics teachers. This is the persistent distance between understanding GenAI in theory and orchestrating it purposefully in classroom practice. This gap cannot be closed by frameworks that simply enumerate competencies. What is missing is a clear account of what purposeful prompt construction looks like across different ped- agogical contexts. Hence, we believe that a new framework must treat prompt literacy not as an ancillary skill but as a foundational dimension of GenAI literacy. 3.6 The Alignment Question: Why a New Framework Is Still Needed Two international frameworks released since 2024 are espe- cially pertinent. UNESCO’s AI Competency Framework for Teachers (Miao & Cukurova, 2024) specifies fifteen competen- cies across five aspects, progressing through three levels (Ac- quire, Deepen, Create); the OECD/European Commission’s Empowering Learners for the Age of AI (2026) organizes AI literacy as a developmental pathway across four domains: En- gage, Create, Manage, and Shape AI. Together they establish an authoritative, human-centered consensus that AI literacy is constituted by ethical and civic capacities, not technical skill alone (cf. Mills et al., 2024). RAIL-Ed’s alignment with them is deliberate, and it is not redundant. These standards specify what teachers and learners should be able to do; they nei- ther theorize the generative-specific dynamics that determine whether a given affordance deepens or displaces learning, nor supply a developmental account of how teachers acquire that capacity. UNESCO’s framework is a competency standard, and the OECD/EU framework is a learner-outcome pathway with limited teacher developmental progression. To make explicit how RAIL-Ed complements these stan- dards, Table 3 maps its six pillars onto the UNESCO AI Com- petency Framework for Teachers (Miao & Cukurova, 2024) and the OECD/EU AI literacy framework (OECD/European Commission, 2026). RAIL-Ed operationalizes their shared commitments for the generative, K–12 teacher-preparation context, adding the developmental progression, the dialectical account of affordances, and the constitutive treatment of equity that those frameworks leave underspecified. 3.7 Rationale for RAIL-Ed Framework Our proposed framework, RAIL-Ed, responds to these con- verging demands. Its six interdependent pillars are not a se- quence of competencies but mutually constituting dimensions of a single literacy orientation, synthesizing Freirean critical pedagogy, Deweyan inquiry, Vygotskian sociocultural the- ory, and Shneiderman’s human-centered design (Dewey, 1938; Freire, 1970/2000; Shneiderman, 2020; Vygotsky, 1978). It is responsive to international standards, the UNESCO AI Com- petency Framework for Teachers (Miao & Cukurova, 2024) and the OECD/EU AI literacy framework (OECD/European Commission, 2026), and to the governance commitments of the OECD AI Principles (OECD, 2024) and the EU AI Act (European Union, 2024). We do not claim to resolve the field’s fragmentation by assertion; we offer a theoretically integrated architecture from which cumulative development of the frame- work can proceed. 4 The RAIL-Ed Framework 4.1 Framework Overview The RAIL-Ed framework synthesizes the gaps documented in Section 3 into an integrative, justice-centered architecture for GenAI literacy across K–12 teacher education, spanning pre-service and in-service teachers. It is built around six in- terdependent pillars (Technical Fluency, Critical Evaluation, Human–AI Collaboration, Contextual Awareness, Ethical Rea- soning, and Empowered Agency). We define GenAI literacy for teachers as the integrated capacity of teachers to understand how generative systems produce outputs, to evaluate those outputs critically, to col- laborate with them in pedagogically and ethically responsible ways, and to exercise principled agency, including restraint, over their role in learning. This definition deliberately extends Long and Magerko’s (2020) canonical account of AI literacy as “a set of competencies that enables individuals to critically evaluate AI technologies” (p. 2) into the generative context, where evaluation must contend with fluent fabrication and collaboration must preserve pedagogical authority. The six pillars operationalize this definition; each is defined below and anchored to the theoretical tradition from which it draws. Technical Fluency is the capacity to understand how genera- tive systems work, how transformer-based models are trained, how reinforcement learning from human feedback shapes their behavior, and why hallucination, bias propagation, and stochasticity are structural, and to translate that understanding into purposeful prompt construction. Grounded in Vygotsky’s (1978) account of mediated cognition, the pillar treats the generative model as a cultural tool that mediates thinking: a teacher who does not grasp how the tool shapes its outputs 8 Table 3: Alignment of RAIL-Ed Pillars With the UNESCO and OECD/EU Frameworks RAIL-Ed pillarsUNESCO AI CFT aspectOECD/EU AILit do- main What RAIL-Ed adds for K–12 teachers Technical Fluency; Critical Evaluation AI foundations & applica- tions Engage with AIGenerative-specific mechanics; prompting as research-design rehearsal; Emerging– Competent–Advanced progression Human–AI Collaboration; Contextual Awareness AI pedagogyCreate with AI; Manage AI Co-agency as a construct; situated, equity- conscious adaptation as a classroom com- petency Ethical Reasoning; Empow- ered Agency Human-centered mindset; Ethics of AI Shape AI Tier-differentiated integrity; principled non-use; dialectical account of risk and benefit Note. UNESCO aspects from Miao and Cukurova (2024); OECD/EU domains from OECD/European Commission (2026). The CFT’s fifth aspect, AI for professional development, is a delivery mechanism spanning all six pillars rather than mapping to any single pillar. RAIL-Ed operationalizes these standards for K–12 teacher preparation; it does not replace them. is mediated by it. It requires teachers to distinguish genera- tive from predictive systems (Annapureddy et al., 2025), to articulate in age-appropriate terms why fluent prose is not evi- dence of truth, and to approach prompting as research-design rehearsal. Critical Evaluation is the ability to critically assess AI out- puts for accuracy, bias, and source credibility. Grounded in Freire’s (1970/2000) critical consciousness and Dewey’s (1938) reflective inquiry, it reframes evaluation from a one- time accuracy check into a habitual, dialogic stance toward AI-mediated knowledge, asking not only whether an output is correct but whose knowledge it centers and whose it omits. It requires teachers to triangulate AI claims against disciplinary sources, to detect fabricated citations and representational bias (Feldman-Maggor et al., 2025), and to model that scrutiny so students internalize it as practice. Human–AI Collaboration is the capacity to engage gen- erative systems as contested co-participants in teaching and learning, negotiating, co-authoring, and iterating with them while preserving pedagogical authority. Grounded in Shnei- derman’s (2020) human-centered AI, which holds that high human control and high automation are independent dimen- sions to be maximized together, the pillar positions the teacher as retaining oversight precisely while leveraging the system’s generativity. Vygotsky’s (1978) mediation extends this to the classroom, where the model functions as a mediational agent in lesson design and inquiry, participating in the work without occupying the social position of a partner (Tate et al., 2026). It requires teachers to plan with AI to sequence scaffold-removal across assignments, and to name and resist voice homogenization (Sourati et al., 2026; Warschauer et al., 2023). Contextual Awareness is the capacity to situate GenAI within the institutional, cultural, linguistic, and political con- ditions in which teaching occurs, recognizing that the same tool produces different consequences across differently re- sourced settings. Grounded in Freire’s (1970/2000) attention to the sociopolitical conditions of knowledge and in sociocul- tural theory’s insistence that cognition is situated, the pillar treats questions of access, power, and whose knowledge is centered as constitutive. It requires teachers to critically exam- ine AI models and data sources, to adapt activities to learners’ contexts, and to anticipate differential affordance outcomes (Cheah et al., 2025). Ethical Reasoning is the capacity to address bias, privacy, authorship, and academic integrity as connected ethical com- mitments. Drawing on Freire (1970/2000), this pillar con- ceptualizes ethics as integral to education. It distinguishes compliance-based, norm-dependent, and principled moral rea- soning, emphasizing teachers’ role in fostering higher levels of moral development. It requires teachers to recognize which tier a learner reasons from, to design conditions that support principled integrity, and to treat transparency about their own AI use: prompt logs, model versions, provenance as responsi- ble epistemic stewardship. Empowered Agency is the capacity to exercise principled, self-authored judgment over GenAI, deciding when to adopt, when to restrain, and when to refuse, and to extend that judg- ment into civic voice and the transformation of AI systems. Grounded in Freire’s (1970/2000) praxis, in which reflection and action combine to transform conditions, and in Shneider- man’s (2020) insistence on sustained human control, the pillar treats principled non-use as a sophisticated literacy posture. It requires teachers to sustain reflective practice under situ- ational pressure, to participate in AI governance rather than accept it passively, and to cultivate the same agency in their students, with participation in design justice as its culminating expression. RAIL-Ed is distinguished from existing models by three architectural commitments. It is integrative, recognizing that the absence of any single pillar creates specific pedagogi- cal failure modes (Table 5). This is the framework’s central testable proposition: if development in one pillar were shown to compensate fully for the absence of another, if, for example, advanced technical fluency alone eliminated the failure modes associated with absent critical evaluation, the integrative com- mitment would be falsified. It is developmental, with each pillar advancing through Emerging, Competent, and Advanced levels (Table 6). It is also dialectical, acknowledging that gen- erative AI affordances simultaneously enable and constrain learning: the same capabilities that support source discov- ery can produce fabricated information (the source discovery– 9 fabrication paradox), and the efficiency gained through AI use can undermine the development of underlying skills. Thus, the outcomes of AI engagement depend on the literacies users bring to these interactions (cf. affordance theory; Gibson, 1979). This dialectical commitment is what allows RAIL-Ed to address the discipline of education’s dual mandate with re- spect to GenAI: not to choose between guarding against its risks and harnessing its potential, but to cultivate the literacies that determine which of the two a shared affordance produces. Generative systems can deepen inquiry, widen access for multi- lingual and underprepared learners, and make expert reasoning visible; the same systems can displace cognition, homogenize voice, and naturalize bias. RAIL-Ed treats these not as a ledger to be balanced but as conditional outcomes of educator and learner capacity, and it specifies the capacities accordingly. In doing so, the framework positions GenAI as a contested epistemic co-participant in the work of teaching and learning rather than as either a threat to be contained or a shortcut to be exploited. Methodologically, the six pillars were derived through a hybrid systematic review (a PRISMA-style search and screen- ing) and qualitative framework analysis, reported in full in a companion review (Zhou et al., in press) and summarized here. From an initial pool of 934 records (ERIC, Web of Science, ScienceDirect, Scopus, IEEE Xplore, and ProQuest; 2023– 2025), 67 studies met the inclusion criteria and were coded against the five analytic dimensions of the concept-analysis matrix: knowledge, skill, ethics, equity, and agency (Walker & Avant, 2005). The six pillars are not a relabeling of these five dimensions but a theoretically motivated restructuring of them into teacher capacities, governed by three criteria: each pillar must name a capacity whose absence produces a distinct, documented pedagogical failure (Table 5); each must admit a distinct developmental trajectory (Table 6); and each must be groundable in the theoretical traditions established in Section 2. Applying these criteria produced two departures from the coding matrix. First, the skill dimension resolved into two pillars, Critical Evaluation and Human–AI Collaboration, because the genera- tive context splits what the corpus treats as one competence: judging outputs that may be fluently fabricated and working with a system as a co-participant are distinct capacities with distinct failure modes (unverified error versus displaced cog- nition). Second, equity was recast from a coverage category into a constitutive commitment: rather than standing as a sep- arate pillar that could be satisfied by topical mention, it is built into Contextual Awareness as the capacity to anticipate differential consequences across differently resourced settings, and it recurs as a design criterion across the remaining pillars. Knowledge maps onto Technical Fluency, ethics onto Ethi- cal Reasoning, and agency onto Empowered Agency; in each case, the literature's content is reframed into the judgment the teacher exercises. Figure 2: The Six Pillars of RAIL-Ed: An Integrative Model of GenAI Literacy for Teachers 4.2 Visual Representation of the Framework Figure 2 renders RAIL-Ed as six interdependent pillars ar- ranged around a central hub representing GenAI literacy for teachers, the integrated capacity defined in Section 4.1. The hub is encircled by a ring carrying the framework's three ar- chitectural commitments, integrative, developmental, and di- alectical, which describe how the pillars work together. Read functionally, and in a grouping distinct from the dimensional mapping developed in Section 4.3, the pillars play three roles: Technical Fluency, Critical Evaluation, and Human–AI Collab- oration supply the cognitive capacities a teacher draws on at the point of decision; Contextual Awareness and Ethical Rea- soning supply the capacity to read the institutional, cultural, and moral conditions within which those decisions are made; and Empowered Agency supplies the disposition that sustains principled practice across changing conditions. The pillars are mutually constituting, which the figure signals by connecting each pillar to the hub and, along the dashed ring, to its neigh- bors. To illustrate: a teacher who notices a fabricated citation in an AI-generated reading list (Critical Evaluation) revises the prompt that produced it (Technical Fluency), decides with the class how the tool should be used (Human–AI Collaboration), weighs what the error means for the particular students in the room (Contextual Awareness and Ethical Reasoning), and models the judgment students are meant to develop (Empow- ered Agency), a single decision moving through all six pillars at once. Table 5 formalizes the converse: the characteristic failures that follow when any one of these capacities is absent. 4.3 Key Components and Dimensions RAIL-Ed organizes GenAI literacy along four interlocking di- mensions: Knowledge, Skill, Ethics, and Critical Dispositions, operationalized through the six pillars. Where Section 4.2 10 grouped the pillars by their function at the point of decision, this section maps them by the kind of capacity each develops; Table 4 provides the crosswalk between each pillar, the gap it addresses, the core competencies and GenAI-specific content it requires, and the pedagogical praxes through which it is enacted. 4.3.1 Knowledge Dimensions AI Basics. Technical Fluency begins from a premise about teaching, not about content: teachers cannot teach or assess what they do not themselves understand (Shulman, 1986). AI4K12’s Five Big Ideas (Touretzky et al., 2019)- perception, representation and reasoning, learning, natural interaction, and societal impact- specify what K–12 students should under- stand about AI; RAIL-Ed does not adopt that scope as its own but treats teacher mastery of it as the floor beneath Technical Fluency. The link is pedagogical, not curricular: AI4K12 defines the student-facing content, while RAIL-Ed defines the teacher capacity required to make that content teachable and to evaluate students’ grasp of it. Two things distinguish the teacher’s capacity from the student’s content. First, RAIL- Ed extends these largely pre-LLM ideas to the distinction AI4K12 never draws between predictive and generative sys- tems (Annapureddy et al., 2025), because the phenomena that matter most for GenAI literacy (hallucination, stochasticity, fluent-but-false output) are properties of generation, not of AI in general. Second, it reframes the Big Ideas from content to be covered into judgment to be exercised: a teacher with Technical Fluency uses this understanding to explain, in age- appropriate terms, why an LLM does not retrieve facts the way a search engine does and why fluent prose is not evidence of truth, and models that reasoning for students. GenAI Mechanics. Educators and learners require func- tional knowledge of how transformer-based LLMs are trained, how reinforcement learning from human feedback (RLHF) shapes their behavior, and why these mechanics make halluci- nation, bias propagation, and stochasticity structural phenom- ena. This pillar responds to Feldman-Maggor et al.’s (2025) demonstration that disciplinary expertise, while necessary for detecting content errors and fabricated references, is not suf- ficient: in their gender-bias example, “the teachers’ TPACK was not enough to critically evaluate the chat output” (p. 7), because recognizing representational bias requires AI-specific knowledge that existing teacher-knowledge frameworks do not capture. RAIL-Ed therefore treats mechanics literacy as a prerequisite for evaluation literacy: a teacher who under- stands why a model generates fluent falsehoods is positioned to evaluate outputs that disciplinary knowledge alone cannot adjudicate. Prompt Engineering. Where existing frameworks treat prompting as ancillary, RAIL-Ed positions it as foundational, consistent with the intervention evidence in Arrington et al. (2025) and with emerging measurement work: Yan et al. (2026) found prompt-crafting so central to competent use that they added it to the usage dimension of their validated GenAI literacy scale, identifying it as a component “previously overlooked in existing scales” (p. 12). Prompt engineering is conceived as research-design rehearsal: the construction of an epistemically front-loaded query in which theoretical com- mitments, audience, and disciplinary conventions are made explicit before the model is engaged. Prompt literacy is nei- ther intuitive nor automatically acquired through tool exposure ( ̧Sim ̧sek, 2025; Wang et al., 2025) and must be taught explic- itly. 4.3.2 The Skill Dimension Educational Planning. Educators must design lessons in which GenAI is integrated as a contested co-participant. This involves selecting tasks where AI augments, not replaces, cognitive labor and sequencing assignments along a scaffold- removal logic that proceeds from AI-permitted brainstorm- ing to independent drafting to closed-environment demonstra- tion. The pillar addresses Yu and Pian’s (2025) documented knowing–doing gap: the persistent distance between under- standing GenAI in theory and orchestrating it purposefully in practice. Assessment Design. Assessment in an LLM-saturated en- vironment requires process-visible architectures that surface how a learner arrived at an outcome, not only what the out- come looks like. Drawing on Bower et al.’s (2024) finding that educators prioritize process-focused assessment in response to generative AI, this competency includes revision logs and reflection journals; tiered tasks distinguishing incidental, sub- stantive, and generative use (Hossain, 2026a); and prompts that LLMs cannot plausibly complete without disciplinary tri- angulation. Consistent with the developmental account of in- tegrity developed below, detection-based assessment is treated as architecturally incapable of producing durable academic integrity (Hossain, 2026a). Communication and Collaboration. RAIL-Ed treats com- munication as both interpersonal and epistemic: communica- tion with students and colleagues about responsible use, and communication with the AI system itself as a documented, citable participant in knowledge production. Voice homoge- nization (Sourati et al., 2026) is treated as a collaboration risk to be named and resisted, especially for multilingual learners (Warschauer et al., 2023). 4.3.3 Ethical Dimensions Bias Awareness. Educators must recognize that training- data composition encodes representational asymmetries (gen- der, racial, linguistic, disciplinary) that surface in classroom- relevant ways: stereotyped examples, narrowed canons, and outputs that perform fluency while obscuring whose knowl- edge has been centered. Every AI output is interrogated for what is present, what is absent, and whose authority the ab- sence reflects (Feldman-Maggor et al., 2025). Privacy and Transparency. Privacy literacy in RAIL-Ed is operational. Educators must know what data the platforms they use collect, what their districts and institutions permit, and what they are obligated to disclose (Miao & Holmes, 2023). Transparency literacy extends to one’s own practice: RAIL-Ed treats prompt logs, model versions, and dataset provenance as responsible epistemic stewardship. 11 Academic Integrity Considerations.Academic in- tegrity in RAIL-Ed is treated developmentally. Drawing on Kohlberg’s (1984) stages of moral development, the frame- work’s three-tier model (Hossain, 2026a) distinguishes three structurally distinct tiers of ethical reasoning, compliance- based (governed by fear of detection), norm-dependent (gov- erned by instructor authority), and principled (governed by self-authored values), that produce different failure modes, such that one-size-fits-all detection policy is architecturally incapable of cultivating principled reasoning. The competency for educators is to recognize which tier a learner is reasoning from and to design conditions that support upward develop- mental movement. 4.3.4 Critical Dispositions Reflective Thinking. Reflective thinking extends Dewey’s (1938) inquiry-based stance into the AI-mediated environment. Educators and learners are asked, recurrently, what intellec- tual work AI performed, displaced, or privileged in a given encounter and what that means for the development of their own judgment. Reflection is not an end-of-term exercise but a structural feature of practice. Ethical Decision-Making. Ethical decision-making is the disposition through which knowledge, skill, and ethical rea- soning are translated into action under situational pressure, a deadline at midnight, a high-stakes assignment, a politicized classroom topic, pressures the prevailing literature has not theorized adequately. Drawing on Freirean praxis (Freire, 1970/2000) and Shneiderman’s (2020) human-centered AI principles, this disposition extends agency beyond individual restraint to include civic voice, participatory critique, and cre- ative transformation of AI systems. RAIL-Ed treats principled non-use as a legitimate outcome of AI literacy rather than a literacy failure: an informed, research-based refusal that constitutes a sophisticated literacy posture (Hossain, 2026a). Design-justice participation is treated as the framework’s cul- minating expression of empowered agency. 4.3.5The Failure-Mode Map: Why Each Pillar Is Load- Bearing RAIL-Ed’s integrative commitment, the claim that the six pil- lars are mutually constituting and jointly necessary, is opera- tionalized in Table 5 as a failure-mode map. Each single-pillar row pairs a capacity with the pedagogical failure documented in the reviewed literature when that capacity is absent; the three compound rows state the framework’s own predictions for what follows when pairs of pillars are missing. The map is the empirical face of the testable proposition stated in Section 4.1. Each row predicts a deficit that development in the remain- ing pillars should not, on the framework’s account, remediate. The catastrophic condition at the foot of the table names the field’s current default, AI-saturated learning without AI liter- acy (Cheah & Kim, 2025), and RAIL-Ed exists to make this default unacceptable. 4.3.6The Developmental Rubric: Emerging, Competent, Advanced Table 6 specifies Emerging, Competent, and Advanced levels for each pillar against which educators, teacher-preparation programs, and accrediting bodies can assess maturation. Two design principles distinguish this rubric from existing devel- opmental schemes. First, advancement in one pillar does not compensate for stagnation in another: an educator who is Advanced in Technical Fluency but Emerging in Ethical Rea- soning produces a compound failure, not a partial success. Second, advancement is recursive: encounters with new AI generations, new student populations, or new policy environ- ments can require recalibration in specific pillars, a feature the rubric treats as healthy practice. 4.4 Application Scenarios Two scenarios, summarized in Table 7, illustrate how the six pillars and developmental rubric translate into concrete classroom decisions. Each is deliberately specific: generic recommendations are the failure mode of prior frameworks, and RAIL-Ed earns its theoretical claims by showing what implementation looks like at the level of an actual unit. 4.4.1 In-Service Teacher Context (K–12 Classroom) In a Grade 7 civics unit, RAIL-Ed shapes instruction at each phase. Technical Fluency is built through a teacher-led prompt walkthrough comparing how three different LLMs respond to the same civics question; the comparison makes tokenization, training-data variation, and probabilistic generation concrete. Critical Evaluation is practiced by auditing an AI summary of a Bill of Rights primary source against the original. Human– AI Collaboration is modeled by the teacher, who plans a class debate aloud with the AI, narrating accept/override/ask-again decisions. Contextual Awareness emerges when the class ex- amines how the AI handles county-level civic specifics and discusses whose knowledge is centered in its training corpus. Ethical Reasoning is introduced through a tier-graduated in- tegrity conversation that asks what the school rule says (Tier 1), what teachers and family expect (Tier 2), and what kind of learner the student wants to be (Tier 3). Empowered Agency is cultivated through a design-justice mini-project, supported by a reflection journal: students propose one classroom rule for AI use and argue for it to the school council. The unit operationalizes the K–12 ethical competencies that Tagare et al. (2026) and Sattelmaier and Pawlowski (2025) identify as central to teaching with AI competencies that Section 3 showed remain marginal in existing literacy frameworks. 4.4.2 Pre-Service Teacher Context (Methods Course) In a pre-service teacher-education methods course, RAIL-Ed is built into the semester’s architecture. Technical Fluency is established through a transformer-mechanics module paired with an iterative prompting task on a primary source, with re- visions logged. Critical Evaluation is operationalized through a verification assignment in which pre-service teachers lo- cate three AI-suggested sources, confirm them in the library database, and reflect on a fabrication they detected. Human– 12 Table 4: RAIL-Ed Pillars: Gaps Addressed, Competencies, GenAI Content, and Pedagogical Praxes RAIL-Ed pillar (dimen- sion) Deficit addressed Core competencies and GenAI- specific content Pedagogical praxes Technical Fluency (Knowledge) Prompt literacy treated as ancil- lary; predictive-vs-GenAI confla- tion (Annapureddy et al., 2025). LLM training and RLHF; halluci- nation and stochasticity; prompt engineering as research-design re- hearsal. Model demonstrations; tool deconstruction; prompt- revision logs. Critical Evaluation (Skill)Frameworks list “evaluation” as a competency without specifying disciplinary application (Sperling et al., 2024). Bias audit; fabricated-citation de- tection; disciplinary triangulation; deepfake analysis. Fact-checking workshops; bias audits; verification pro- tocols. Human–AI Collaboration (Skill) Agency narrowed to tool-adoption intention; human-in-the-loop under-operationalized (Su & Yang, 2023). Ethical co-engagement; co- authorship; prompt iteration; GenAI-assisted lesson design. Case-based PD; design chal- lenges; scaffold-removal sequencing. Contextual Awareness (Knowledge–Skill bridge) Frameworks decontextualized from K–12 vs. HE realities; weak attention to power and language (Cheah & Kim, 2025). Political economy of foundation models; data colonialism; AI gover- nance; linguistic adaptation. Sociocultural inquiry; policy analysis; multilingual com- parison. Ethical Reasoning (Ethics) Ethics treated as compliance checklist; bias, privacy, integrity addressed in isolation (Tagare et al., 2026). Authorship and surveillance; LLM training-data privacy; tier- differentiated integrity. Ethical scenarios; phronesis- based reflection; develop- mental disclosure. Empowered Agency (Crit- ical Dispositions) Affective and civic dimensions absent or reduced to technology acceptance. Civic voice; participatory critique; creative transformation; principled non-use. Participatory design; advo- cacy projects; design-justice activities. Note. Gaps are documented in the systematic review reported in Section 3. Pedagogical praxes are illustrative. PD = professional development. Table 5: Failure-Mode Map: Documented Consequences of Absent or Underdeveloped Pillars Absent or underdeveloped capacityResulting pedagogical failure mode Technical FluencyUsers conflate LLMs with search engines and read fluent prose as evidence of truth, leav- ing hallucination unanticipated (the LLM-as-oracle error). Critical EvaluationFabricated citations, biased framings, and unverified claims pass unchecked into student and teacher work. Human–AI CollaborationAI is either banned outright or used as an unsupervised substitute; cognitive labor is dis- placed rather than scaffolded. Contextual AwarenessActivities imported from well-resourced settings deepen access and language asymmetries; questions of whose knowledge is centered go unasked. Ethical ReasoningIntegrity collapses into detection; learners comply under surveillance but internalize no principle (compliance theater). Empowered Agency Adoption proceeds by peer default or reflexive refusal; no durable, self-authored stance survives changing conditions. Technical Fluency + Critical Evaluation (compound) Fluent dependence: confident, well-prompted use of output that is never verified. Human–AI Collaboration + Critical Eval- uation (compound) Confident error: smoothly integrated work built on fabricated or biased material. Critical Evaluation + Technical Fluency / prompt literacy (compound) Degraded scrutiny: careful checking of outputs already compromised by the questions that produced them. All pillars absent (catastrophic) AI-saturated learning without AI literacy—the field’s current default condition (Cheah & Kim, 2025). Note. Single-pillar failure modes synthesize gaps documented in the Section 3 review; compound failure modes are the framework’s predictions, stated as testable claims. 13 Table 6: Developmental Rubric for the Six Pillars of RAIL-Ed PillarEmergingCompetentAdvanced Technical FluencyUses default prompts; conflates LLMs with search engines; lim- ited awareness of hallucination. Designs purposeful prompts; ex- plains transformer mechanics and RLHF; recognizes hallucination as structural. Builds replicable prompting work- flows; teaches mechanics through guided discovery; publishes prompt libraries. Critical Evaluation Accepts AI outputs at face value; rarely cross-checks citations; treats fluent prose as accurate. Triangulates against scholarly databases; detects fabrications and bias routinely; redesigns assess- ments. Theorizes the source discovery– fabrication paradox in disciplinary terms; mentors students into verifica- tion. Human–AI Collabo- ration Treats AI as prohibited or as an unsupervised substitute; planning happens before or after AI, not with it. Plans lessons with AI as a con- tested co-participant; sequences scaffold-removal assignments; models negotiation aloud. Orchestrates multi-tool workflows; designs assignments exploiting AI while building independent compe- tence. Contextual Aware- ness Imports activities from elite- institution contexts without adap- tation; unaware of access asym- metries. Adapts activities to learner back- ground, language, and context; recognizes differential affordance outcomes. Designs equity-conscious AI cur- ricula at program or institutional level; contributes to MSI- or region- specific adaptations. Ethical ReasoningTreats integrity as detection; relies on punitive policy; cannot distinguish compliance from principled reasoning. Uses tier-differentiated citation pedagogy; distinguishes inciden- tal, substantive, and generative use. Develops institutional AI gover- nance grounded in developmental ethics; treats first-time errors as ad- vancement opportunities. Empowered AgencyAdopts AI because peers do; cannot articulate when to restrain or refuse; reflection is episodic. Maintains reflective practice; models situated decision-making; recognizes principled non-use as legitimate. Sustains principled (Tier 3) reason- ing under pressure; participates in AI governance; cultivates civic critique in learners. Note. Advancement in one pillar does not compensate for stagnation in another; advancement is recursive. MSI = minority-serving institution. AI Collaboration is enacted through a lesson-plan co-drafting exercise: candidates generate a lesson plan with GenAI, revise it independently, and submit both versions plus a reflection on what they changed and why. Contextual Awareness is taught through a multilingual comparison exercise examining whose voice AI editing preserves and whose it smooths away (Warschauer et al., 2023). Ethical Reasoning is structured by a course-wide AI citation protocol distinguishing incidental, substantive, and generative use, with first-time disclosure er- rors treated as developmental occasions. Empowered Agency is cultivated through an end-of-term portfolio that includes a Reliance Negotiation Statement (Hossain, 2026b), a reflec- tive account in which candidates describe when and why they used, restrained, or refused GenAI across the semester, and a public-facing op-ed advocating an institutional AI policy position. These scenarios demonstrate that RAIL-Ed accommo- dates the maturation of each pillar across the K–12 teacher- preparation continuum, from pre-service preparation through in-service practice. A teacher candidate practicing Tier 1 and Tier 2 ethical reasoning in a methods course is building the scaffolding from which Tier 3 principled reasoning matures across years of classroom practice. The framework’s integra- tive, developmental, and dialectical commitments distinguish RAIL-Ed from the additive competency lists the field has, to date, produced. 5Implications for Practice, Policy, and Re- search The RAIL-Ed framework aims to reposition GenAI literacy as more than functional knowledge or skills in using classroom tools. Earlier AI literacy frameworks established important foundations for understanding, using, and evaluating AI (Long & Magerko, 2020; Ng et al., 2021), but the rise of GenAI introduces new epistemic, pedagogical, and ethical demands: teachers must now judge fluent but fallible outputs, negotiate human-AI authorship, identify bias embedded across the AI life cycle, and decide when AI supports or suppresses learning (Kasneci et al., 2023; Lee et al., 2024). The significance of RAIL-Ed therefore lies in its shift from competency acqui- sition to educational judgment. It treats responsible GenAI literacy as an institutional and pedagogical project: teacher preparation and training curricula must cultivate verification and disciplinary reasoning, teacher learning must develop sit- uated professional judgment, policy must protect equity and accountability, and research must validate evidence of actual learning. 5.1 Implications for Practice The RAIL-Ed framework implies that GenAI literacy should reshape how teacher preparation and training are designed. The central pedagogical implication is that teachers need to shift from output-centered instruction toward pedagogies that make reasoning, verification, revision, and judgment visible. Teachers’ responsible GenAI literacy is treated as the intended outcome of this preparation, not simply their technical fluency 14 Table 7: RAIL-Ed Pillars Applied to In-Service and Pre-Service K–12 Teacher Contexts PillarK–12 application (Grade 7 civics unit)Pre-service teacher application (methods course) Technical FluencyTeacher-led prompt walkthrough comparing how three LLMs answer the same civics question; tok- enization made concrete. Short transformer-mechanics module paired with it- erative prompting on a primary-source task; prompt revisions logged. Critical EvaluationPupils audit AI summaries of a Bill of Rights pri- mary source against the original, identifying omis- sions and tone shifts. Verification assignment: candidates locate three AI-suggested sources, confirm them in the library database, and reflect on a fabrication they detected. Human–AI Collabora- tion Teacher plans a class debate aloud with AI, narrating accept/override/ask-again decisions students will themselves make. Lesson-plan co-drafting: candidates generate a lesson plan with GenAI, revise it independently, and submit both versions plus a reflection on their changes. Contextual Awareness Class examines how AI handles county-level civic specifics it is unlikely to represent well; discusses whose knowledge is centered. Candidates run a multilingual comparison of AI- edited drafts in home language and English, examin- ing whose voice AI preserves or smooths away. Ethical ReasoningTier-graduated integrity conversation: what the rule says (Tier 1), what teachers and family expect (Tier 2), who you want to be (Tier 3). Course-wide citation protocol distinguishing inci- dental, substantive, and generative use; first-time errors as developmental occasions. Empowered AgencyReflection journal plus a design-justice mini-project: students propose one classroom AI rule and argue for it to the school council. End-of-term portfolio with a Reliance Negotiation Statement and a public-facing op-ed on institutional AI policy. Note. Scenarios are illustrative implementations of the framework, not prescribed curricula; local adaptation is expected. with a tool. Teacher candidates and in-service teachers learn not only how to use GenAI, but also how to question a source, check evidence, and notice when fluent language hides weak reasoning or bias, so that they are equipped to model and scaffold this judgment for the students they teach (Kasneci et al., 2023). This shift requires embedding GenAI literacy within dis- ciplinary inquiry. For example, in science, teachers learn to examine how an AI explanation simplifies causal mechanisms or misrepresents uncertainty, and to design lessons that sur- face those simplifications for their students. In social studies methods courses, teachers practice comparing AI-generated claims with primary sources to detect hallucinations while ask- ing whose perspectives are presented or missed, then translate that practice into classroom tasks. These pedagogical moves position GenAI, for teachers, beyond a tool for productivity to an object of critique and inquiry that they must first master before they can teach it. The same shift appears in assessment, which now captures teachers’ own epistemic activity as GenAI users. A polished lesson plan, unit design, or written reflection submitted in a methods course or professional-development module no longer gives teacher educators the same evidence it once did. A teacher candidate may have written it independently, re- vised AI output carefully, or accepted generated text with little understanding of the pedagogical reasoning behind it. The final product alone may not tell the difference. For this rea- son, teachers need evidence of how they arrived at a lesson, assessment, or communication: what they accepted, rejected, verified, or questioned. RAIL-Ed therefore treats responsi- ble GenAI use as evidence that teachers really develop this judgment instead of a rule they must obey. Finally, pedagogical practice should cultivate discipline as well as use in both teaching and learning. RAIL-Ed includes knowing when not to use AI, when to slow down, when to seek human feedback, and when AI may flatten uncertainty, voice, culture, or disciplinary complexity. Teachers do not need to become computer scientists, but they do need a basic understanding of how GenAI works. They should know, for example, that LLMs generate likely sequences of words, not verified facts. This helps explain why a model can produce a fluent answer, a false citation, or a confident explanation that is only partly correct (Bender et al., 2021). Without this understanding, teachers may trust AI outputs too quickly. They may also reject technology entirely and miss situations where it could support learning. The practical implication of RAIL-Ed is that teachers should learn to design learning environments with responsible AI use so that students consistently practice disciplined decisions on AI’s role in learning. 5.2 Implications for Policy and Governance RAIL-Ed suggests that school and institutional policy move beyond reactive control of GenAI use toward governance that supports responsible and authentic learning. Many current policies respond to GenAI by focusing on plagiarism detec- tion, disclosure rules, or bans on AI-assisted work. These responses are understandable, with concerns about authorship and academic integrity. However, they are insufficient as a policy foundation because they treat GenAI primarily as a threat to academic integrity. RAIL-Ed-informed stakeholders would ask whether institutional rules help teachers develop judgment, accountability, and agency in AI-assisted work. Policy should therefore provide clearer guidance across different professional situations. A teacher who uses AI to brainstorm lesson ideas is not doing the same thing as a teacher who submits a fully AI-generated lesson plan, assessment, or parent communication without review. AI-assisted editing, 15 resource curation, translation support, and co-planning also raise different pedagogical and ethical questions. Institutions need tiered guidance that names these differences and connects them to professional purposes: when AI use is discouraged, when it is permitted with disclosure, when it is scaffolded as part of professional development, and when it is encouraged for critical evaluation or accessibility. Otherwise, teacher- preparation programs and school leaders may be left to guess what counts as acceptable use. Governance must address due process and equity. Policy must account for unequal access to paid AI tools, reliable devices, teacher support, and dominant-language advantage. Without explicit equity provisions, GenAI may widen existing educational inequalities. Finally, RAIL-Ed frames policy as an institutional responsibility, not a classroom management issue. Institutions should establish standards for privacy, data protection, accessibility, teacher professional learning, and human oversight. This aligns with broader international guid- ance emphasizing human agency, inclusion, equity, privacy, transparency, and accountability in AI governance (OECD, 2024; Miao & Holmes, 2023). 5.3 Implications for Future Research RAIL-Ed also has implications for how GenAI literacy should be studied. Instead of relying mainly on surveys or accep- tance models, researchers should track evidence of AI-assisted learning. This includes how learners verify information, re- vise AI-generated text, explain their choices, and decide when to use AI and when not. Longitudinal and classroom-based studies would be especially useful because GenAI literacy develops through repeated use, feedback, mistakes, and chang- ing expectations. Such evidence would allow researchers to distinguish between superficial AI use and deeper forms of critical evaluation, ethical reasoning, and empowered agency. RAIL-Ed also encourages researchers to study GenAI liter- acy as a situated educational practice. Teachers’ responsible AI use depends on preparation-program design, mentorship and coaching, institutional policy, tool access, and disciplinary norms. Research should therefore attend to the professional en- vironment in which AI literacy is enacted. Previous research has found that AI systems can intensify existing inequities through unequal access, dominant-language advantage, and biased outputs across the AI life cycle (Lee et al., 2024). The research implication of RAIL-Ed is methodological as well as conceptual, and studies should examine how human judgment develops within socio-technical systems. RAIL-Ed suggests that the field needs stronger alignment between constructs and evidence. When AI literacy is defined as technical fluency, crit- ical evaluation, human-AI collaboration, contextual awareness, ethical reasoning, and empowered agency, research instru- ments and analytic methods must capture these dimensions directly. Otherwise, the field risks measuring convenience constructs, such as frequency of use or perceived usefulness, while claiming to study literacy. A RAIL-Ed-informed re- search agenda would therefore push AI literacy scholarship to- ward richer evidence of teacher learning and stronger construct validity where responsible AI judgment becomes possible. 6 Limitations and Directions for Future Work 6.1 Conceptual Scope and Boundaries The RAIL-Ed framework is conceptually oriented and nec- essarily bounded in scope. It synthesizes critical pedagogy, pragmatist inquiry, sociocultural theory, and human-centered AI design, yet it does not exhaust the theoretical resources available for thinking about GenAI in education (Long & Magerko, 2020; Ng et al., 2021). Adjacent traditions, in- cluding posthumanist and feminist epistemologies, decolonial perspectives, and disability studies in education, could extend the framework in productive ways and surface tensions that the present synthesis addresses only partially (Nemorin, 2024). The six pillars operate at a level of abstraction that requires in- terpretive translation into specific disciplines, grade levels, and institutional cultures (Ng et al., 2021; Zhang et al., 2025). The framework does not prescribe particular instructional meth- ods or assessment instruments, nor does it claim universal applicability across educational settings. It offers, rather, a coherent scaffold that scholars and practitioners can adapt, contest, and refine in light of context-specific commitments and constraints. 6.2 Areas for Empirical Research A further limitation concerns the framework's evidentiary base. Several of the constructs that inform RAIL-Ed, including the three-tier ethical reasoning model developed in Section 4.3 and the reliance-negotiation account behind its integrity pedagogy, derive from the lead author's recent mixed-methods research (Hossain, 2026a, 2026b), conducted at a single minority- serving institution and not yet independently replicated or peer reviewed. Independent testing across institutions, popula- tions, and grade levels is a precondition for the framework's broader claims, and we flag this dependency so adopters can weigh its propositions accordingly. This need is not unique to RAIL-Ed; recent umbrella re- views show that empirical work on AI literacy remains uneven across grade levels, methodological traditions, and disciplinary contexts (Fu et al., 2025). Two directions follow specifically from the framework's internal structure, complementing the broader agenda in Section 5.3. First, studies might examine relationships among the pillars, such as how technical flu- ency mediates ethical reasoning or how contextual awareness shapes human-AI collaboration (Lee et al., 2024). Second, design-based research could refine curricular and instructional approaches grounded in these interdependencies, clarifying which pillars require domain-specific elaboration and which travel across content areas. 6.3 Potential for Cross-National Adaptation Although the framework draws on broadly applicable theoret- ical traditions, its instantiation will vary considerably across national, cultural, and policy contexts. Educational systems differ in their governance structures, infrastructural capacities, linguistic ecologies, and prevailing pedagogical philosophies, and these differences shape what responsible AI literacy can and should mean in particular settings (OECD, 2024; Miao 16 & Holmes, 2023). The framework therefore invites cross- national adaptation. Scholars working in the Global South have noted that frameworks developed primarily in North American and European contexts often underrepresent local epistemologies, languages, and infrastructural realities, and that the ethics of AI in education itself reflects Eurocentric assumptions that warrant decolonial interrogation (Nemorin, 2024). Comparative research that situates the RAIL-Ed frame- work within emerging regional policy environments, including the African Union's (2024) Continental AI Strategy and paral- lel initiatives across ASEAN and Latin American educational ministries, could surface productive tensions and reveal op- portunities for context-sensitive refinement. Such work would strengthen the framework's global relevance and its capacity to support equitable, culturally grounded responses to GenAI in education. 7 Conclusion This paper has proposed the RAIL-Ed framework as a theo- retically integrated response to the conceptual fragmentation in current GenAI literacy discourse (Long & Magerko, 2020; Ng et al., 2021). The framework articulates six interdepen- dent pillars: technical fluency, critical evaluation, human-AI collaboration, contextual awareness, ethical reasoning, and em- powered agency. Drawing on Freire's (1970/2000) pedagogy of critical consciousness, Dewey's (1938) reflective inquiry, Vygotsky's (1978) sociocultural theory of mediated cogni- tion, and Shneiderman's (2020) principles of human-centered AI, RAIL-Ed positions GenAI literacy as a relational, situ- ated, and ethically grounded practice. The framework offers a coherent scaffold for curriculum design, teacher education, professional development, and policy formation across K–12 contexts (OECD, 2024; Miao & Holmes, 2023). Central to this position is RAIL-Ed's rejection of deficit- oriented approaches, which attribute educational challenges to perceived inadequacies in learners or communities rather than to the structural conditions shaping access and participation (Davis & Museus, 2019; Tewell, 2020). In GenAI contexts, deficit framing recasts students' struggles as insufficient digital readiness and narrows literacy to prompt compliance, obscur- ing the algorithmic and institutional conditions that produce those struggles (Kay et al., 2024; Lee et al., 2024). Freire's (1970/2000) critique of the banking model offers a corrective: learners are not vessels to be filled with AI skills but active participants capable of interrogating the systems that mediate their knowledge. The responsible integration of GenAI in education is not a technical problem to be solved but an ongoing pedagogical, ethical, and political project. Researchers should test, contest, and extend the framework through studies attentive to context, culture, and equity, partic- ularly in settings underrepresented in the literature (Nemorin, 2024; Zhang et al., 2025). Educators can treat it not as a prescription but as a generative scaffold for designing experi- ences in which students question GenAI, collaborate with it critically, and shape its place in their learning (Selwyn, 2022). Policymakers, in turn, should align curricular, infrastructural, and regulatory commitments with the principles of human oversight, equity, and justice articulated in international and regional guidance (African Union, 2024; OECD, 2024; Miao & Holmes, 2023). Building a responsible GenAI literacy is necessarily collective, requiring sustained collaboration across disciplines, sectors, and geographies and a willingness to re- frame AI literacy as a shared commitment to human flourish- ing in an algorithmic age. References African Union. (2024). Continental artificial intelligence strategy: Harnessing AI for Africa's development and pros- perity. https://au.int/sites/default/files/documents/44004 -doc-EN-_Continental_AI_Strategy_July_2024.pdf Allen, L. K., & Kendeou, P. (2024). ED-AI Lit: An interdis- ciplinary framework for AI literacy in education. Policy Insights from the Behavioral and Brain Sciences, 11(1), 3–10. https://doi.org/10.1177/23727322231220339 Annapureddy, R., Fornaroli, A., & Gatica-Perez, D. (2025). Generative AI literacy: Twelve defining competencies. Digital Government: Research and Practice, 6(1), Article 13, 1–21. https://doi.org/10.1145/3685680 Arrington, C., McVey, D., Mativo, J., & Pidaparti, R. M. (2025). Enhancing teacher effectiveness with AI-based prompt engineering: A proof of concept. Journal of STEM Education: Innovations and Research, 26(2), 5–10. https: //doi.org/10.63504/jstem.v26i2.2711 Bae, H., Hur, J., Park, J., Choi, G. W., & Moon, J. (2024). Pre-service teachers' dual perspectives on generative AI: Benefits, challenges, and integrating into teaching and learning. Online Learning, 28(3), 131–156. https://doi.or g/10.24059/olj.v28i3.4543 Barbieri, W., & Nguyen, N. (2025). Generative AI as a "place- ment buddy": Supporting pre-service teachers in work- integrated learning, self-management and crisis resolution. Australasian Journal of Educational Technology, 41(2), 34–49. https://doi.org/10.14742/ajet.10035 Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can lan- guage models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (p. 610–623). Association for Computing Machinery. https://doi.org/10.1145/3442188.3445922 Bower, M., Torrington, J., Lai, J. W. M., Petocz, P., & Al- fano, M. (2024). How should we change teaching and assessment in response to increasingly powerful gener- ative artificial intelligence? Outcomes of the ChatGPT teacher survey. Education and Information Technologies, 29(12), 15403–15439. https://doi.org/10.1007/s10639-0 23-12405-0 Bukar, U. A., Sayeed, M. S., Razak, S. F. A., Yogarayan, S., & Sneesl, R. (2024). Prioritizing ethical conundrums in the utilization of ChatGPT in education through an analytical hierarchical approach. Education Sciences, 14(9), Article 959. https://doi.org/10.3390/educsci14090959 17 Cheah, Y. H., Lu, J., & Kim, J. (2025). Integrating genera- tive artificial intelligence in K–12 education: Examining teachers' preparedness, practices, and barriers. Computers and Education: Artificial Intelligence, 8, Article 100363. https://doi.org/10.1016/j.caeai.2025.100363 Chiu, T. K. F., Ahmad, Z., Ismailov, M., & Sanusi, I. T. (2024). What are artificial intelligence literacy and competency? A comprehensive framework to support them. Computers and Education Open, 6, Article 100171. https://doi.org/10 .1016/j.caeo.2024.100171 Choi, Y. S. (2025). Earth science simulations with genera- tive artificial intelligence (GenAI). Journal of University Teaching and Learning Practice, 22(1). https://doi.org/10.53761/nf1yqr46 Davis, L. P., & Museus, S. D. (2019). What is deficit thinking? An analysis of conceptualizations of deficit thinking and implications for scholarly research. Currents: Journal of Diversity Scholarship for Social Change, 1(1). https: //doi.org/10.3998/currents.17387731.0001.110 Dewey, J. (1938). Experience and education. Macmillan. European Union. (2024). REGULATION (EU) 2024/1689 OF THE EUROPEAN PARLIAMENT AND OF THE COUNCIL of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). In eur-lex.europa.eu. Official Journal of the European Union. https://eur-lex.europa.eu/eli/reg/2024/1689/oj Feldman-Maggor, Y., Blonder, R., & Alexandron, G. (2025). Perspectives of generative AI in chemistry education within the TPACK framework. Journal of Science Ed- ucation and Technology, 34(1), 1–12. https://doi.org/10.1 007/s10956-024-10147-3 Freire, P. (2000). Pedagogy of the oppressed (M. B. Ramos, Trans.; 30th anniversary ed.). Continuum. (Original work published 1970) Fu, Y., Weng, Z., & Wang, J. (2025). Examining AI use in edu- cational contexts: A scoping meta-review and bibliometric analysis. International Journal of Artificial Intelligence in Education, 35(3), 1388–1444. https://doi.org/10.1007/s4 0593-024-00442-w Gibson, J. J. (1979). The ecological approach to visual per- ception. Houghton Mifflin. Gruenhagen, J. H., Sinclair, P. M., Carroll, J.-A., Baker, P. R. A., Wilson, A., & Demant, D. (2024). The rapid rise of generative AI and its implications for academic integrity: Students' perceptions and use of chatbots for assistance with assessments. Computers and Education: Artificial Intelligence, 7, Article 100273. https://doi.org/10.1016/j. caeai.2024.100273 Hossain, S. (2026a). Measuring college students’ reliance on large language models in academic writing: A sequential explanatory mixed-methods study at a minority-serving institution (Doctoral dissertation, University of Maryland, Baltimore County). Hossain, S. (2026b). The reliance negotiation framework: A dynamic process model of student LLM engagement in academic writing. arXiv preprint arXiv:2604.16772. Jin, Y., Martinez-Maldonado, R., Gaševi ́ c, D., & Yan, L. (2025). GLAT: The generative AI literacy assessment test. Computers and Education: Artificial Intelligence, 9, Arti- cle 100436. https://doi.org/10.1016/j.caeai.2025.100436 Jorolan, J., Cabillo, F., Batucan, R. R., Camansi, C. M., Eto- quilla, A., Gapo, J., Hallarte, D. K., Milano, M. L., Gonza- les, R., & Gonzales, G. (2025). Development and valida- tion of moral ascendancy and dependency in AI integration (MAD-AI) scale for teachers. Computers & Education, 235, Article 105346. https://doi.org/10.1016/j.compedu. 2025.105346 Joseph, S. (2023). Large language model-based tools in language teaching to develop critical thinking and sus- tainable cognitive structures. Rupkatha Journal, 15(4). https://doi.org/10.21659/rupkatha.v15n4.13 Kasneci, E., Sessler, K., Küchemann, S., Bannert, M., Demen- tieva, D., Fischer, F., Gasser, U., Groh, G., Günnemann, S., Hüllermeier, E., Krusche, S., Kutyniok, G., Michaeli, T., Nerdel, C., Pfeffer, J., Poquet, O., Sailer, M., Schmidt, A., Seidel, T., . . . Kasneci, G. (2023). ChatGPT for good? On opportunities and challenges of large language models for education. Learning and Individual Differences, 103, 102274. https://doi.org/10.1016/j.lindif.2023.102274 Kay, J., Kasirzadeh, A., & Mohamed, S. (2024). Epistemic injustice in generative AI. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 7(1), 684–697. http s://doi.org/10.1609/aies.v7i1.31671 Kohlberg, L. (1984). The psychology of moral development: The nature and validity of moral stages. Harper & Row. Kong, S. C., Yang, Y., & Hou, C. (2024). Examining teach- ers’ behavioural intention of using generative artificial intelligence tools for teaching and learning based on the extended technology acceptance model. Computers and Education: Artificial Intelligence, 7, Article 100328. https://doi.org/10.1016/j.caeai.2024.100328 Laine, J., Minkkinen, M., & Mäntymäki, M. (2025). Under- standing the ethics of generative AI: Established and new ethical principles. Communications of the Association for Information Systems, 56, 1–25. https://doi.org/10.17705/1 CAIS.05601 Lee, J., Hicke, Y., Yu, R., Brooks, C., & Kizilcec, R. F. (2024). The life cycle of large language models in ed- ucation: A framework for understanding sources of bias. British Journal of Educational Technology, 55(5), 1982– 2002. https://doi.org/10.1111/bjet.13505 Long, D., & Magerko, B. (2020). What is AI literacy? Com- petencies and design considerations. In Proceedings of the 2020 CHI Conference on Human Factors in Comput- ing Systems, 1–16. Association for Computing Machinery. https://doi.org/10.1145/3313831.3376727 MacDowell, P., Moskalyk, K., Korchinski, K., & Morrison, D. (2024). Preparing educators to teach and create with 18 generative artificial intelligence. Canadian Journal of Learning and Technology, 50(4), 1–23. https://doi.org/10 .21432/cjlt28606 Memarian, B., & Doleck, T. (2024). Human-in-the-loop in artificial intelligence in education: A review and entity- relationship (ER) analysis. Computers in Human Behavior Artificial Humans, 2(1), Article 100053. https://doi.org/10 .1016/j.chbah.2024.100053 Miao, F., & Cukurova, M. (2024). AI competency framework for teachers. UNESCO. https://doi.org/10.54675/ZJTE208 4 Miao, F., & Holmes, W. (2023). Guidance for Generative AI in Education and Research. UNESCO Publishing. https: //unesdoc.unesco.org/ark:/48223/pf0000386693 Miao, F., Holmes, W., Huang, R., & Zhang, H. (2021). AI and education: Guidance for policy-makers. UNESCO. https://doi.org/10.54675/pcsp7350 Miao, F., Shiohira, K., & Lao, N. (2024). AI competency framework for students. UNESCO. https://doi.org/10.546 75/JKJB9835 Mills, K., Ruiz, P., Lee, K., Coenraad, M., Fusco, J., Roschelle, J., & Weisgrau, J. (2024). AI literacy: A framework to un- derstand, evaluate, and use emerging technology. Digital Promise. https://doi.org/10.51388/20.500.12265/218 Mollick, E. R., & Mollick, L. (2023). Assigning AI: Seven approaches for students, with prompts (Preprint). arXiv. https://doi.org/10.48550/arXiv.2306.10052 Nemorin, S. (2024). Towards decolonising the ethics of AI in education. Globalisation, Societies and Education, 22(1), 1260–1272. https://doi.org/10.1080/14767724.2024.2333 821 Ng, D. T. K., Leung, J. K. L., Chu, S. K. W., & Qiao, M. S. (2021). Conceptualizing AI literacy: An exploratory review. Computers and Education: Artificial Intelligence, 2, Article 100041. https://doi.org/10.1016/j.caeai.2021.1 00041 Ng, D. T. K., Chan, E. K. C., & Lo, C. K. (2025). Op- portunities, challenges and school strategies for integrat- ing generative AI in education. Computers and Educa- tion: Artificial Intelligence, 8, Article 100373. https: //doi.org/10.1016/j.caeai.2025.100373 OECD. (2019). OECD future of education and skills 2030: OECD learning compass 2030. A series of concept notes. OECD Publishing. https://w.oecd.org/content/dam/oec d/en/about/projects/edu/education-2040/1-1-learning-c ompass/OECD_Learning_Compass_2030_Concept_No te_Series.pdf OECD. (2024). The OECD Artificial Intelligence (AI) Princi- ples. In oecd.ai. OECD. https://oecd.ai/en/ai-principles OECD/European Commission. (2026). Empowering learners for the age of AI: An AI literacy framework for primary and secondary education. OECD Publishing. https://doi. org/10.1787/65cd27d4-en Ofosu-Asare, Y. (2025). Guiding the future: Developing an ethical framework for generative AI use in education. Inter- national Journal of Information and Learning Technology, 43(1), 26 – 44. https://doi.org/10.1108/IJILT-06-2024-011 3 Park, J. (2025). A systematic literature review of generative artificial intelligence (GenAI) literacy in schools. Com- puters and Education: Artificial Intelligence, 9, Article 100487. https://doi.org/10.1016/j.caeai.2025.100487 Roe, J., Perkins, M., & Furze, L. (2025). Reflecting reality, amplifying bias? Using metaphors to teach critical AI lit- eracy. Journal of Interactive Media in Education, 2025(1), Article 18. https://doi.org/10.5334/jime.961 Sattelmaier, L., & Pawlowski, J. (2025). Be AIware! An AI competency model for K–12 education. Social Sciences & Humanities Open, 12, Article 101838. https://doi.org/10.1 016/j.ssaho.2025.101838 Selwyn, N. (2022). The future of AI and education: Some cautionary notes. European Journal of Education, 57(4), 620–631. https://doi.org/10.1111/ejed.12532 Shneiderman, B. (2020). Human-centered artificial intelli- gence: Reliable, safe & trustworthy. International Jour- nal of Human–Computer Interaction, 36(6), 495–504. https://doi.org/10.1080/10447318.2020.1741118 Shulman, L. S. (1986). Those who understand: Knowledge growth in teaching. Educational Researcher, 15(2), 4–14. https://doi.org/10.3102/0013189X015002004 ̧Sim ̧sek, N. (2025). Integration of ChatGPT in mathemati- cal story-focused 5E lesson planning: Teachers and pre- service teachers' interactions with ChatGPT. Education and Information Technologies, 30, 11391–11462. https: //doi.org/10.1007/s10639-024-13258-x Sourati, Z., Ziabari, A. S., & Dehghani, M. (2026). The homogenizing effect of large language models on human expression and thought. Trends in Cognitive Sciences. Advance online publication. https://doi.org/10.1016/j.tics .2026.01.003 Sperling, K., Stenberg, C. J., McGrath, C., Åkerfeldt, A., Heintz, F., & Stenliden, L. (2024). In search of artificial intelligence (AI) literacy in teacher education: A scoping review. Computers and Education Open, 6, Article 100169. https://doi.org/10.1016/j.caeo.2024.100169 Su, J., & Yang, W. (2023). Unlocking the power of Chat- GPT: A framework for applying generative AI in ed- ucation. ECNU Review of Education, 6(3), 355–366. https://doi.org/10.1177/20965311231168423 Tagare, D., Karki, T., & Yu, W. (2025). K–12 teachers' ethical competencies for AI literacy: Insights from a systematic literature review. Computers & Education, 239, Article 105435. https://doi.org/10.1016/j.compedu.2025.105435 Tan, Q. (2025). Reimagining teacher development in the era of generative AI: A scoping review. Teaching and Teacher Education, 168, Article 105236. https://doi.org/10.1016/j. tate.2025.105236 19 Tate, T. P., Ritchie, D. R., & Warschauer, M. (2026). Gen- erative AI as a mediational agent: Rethinking learn- ing in sociocultural theory (Preprint). EdArXiv. https: //doi.org/10.35542/osf.io/wcpj5_v1 Tewell, E. (2020). The problem with grit: Dismantling deficit thinking in library instruction. portal: Libraries and the Academy, 20(1), 137–159. https://doi.org/10.1353/pla.20 20.0007 Thomas, J., & Harden, A. (2008). Methods for the thematic synthesis of qualitative research in systematic reviews. BMC Medical Research Methodology, 8, Article 45. https: //doi.org/10.1186/1471-2288-8-45 Touretzky, D., Gardner-McCune, C., Martin, F., & Seehorn, D. (2019). Envisioning AI for K–12: What should every child know about AI? Proceedings of the AAAI Conference on Artificial Intelligence, 33(1), 9795–9799. https://doi.org/ 10.1609/aaai.v33i01.33019795 Vygotsky, L. S. (1978). Mind in society: The development of higher psychological processes (M. Cole, V. John-Steiner, S. Scribner, & E. Souberman, Eds.). Harvard University Press. Walker, L. O., & Avant, K. C. (2005). Strategies for theory construction in nursing (4th ed.). Pearson Prentice Hall. Wang, Y., Zhang, T., Yao, L., & Seedhouse, P. (2025). A scoping review of empirical studies on generative artificial intelligence in language education. Innovation in Lan- guage Learning and Teaching, 1-28. https://doi.org/10.1 080/17501229.2025.2509759 Warschauer, M., Tseng, W., Yim, S., Webster, T., Jacob, S., Du, Q., & Tate, T. (2023). The affordances and contradictions of AI-generated text for writers of English as a second or foreign language. Journal of Second Language Writing, 62, Article 101071. https://doi.org/10.1016/j.jslw.2023.10 1071 Yan, W., Liu, Y.-l., Mamaeva, V., Dong, F., Tao, G., Li, R., & Yang, H. (2026). Generative AI literacy: Scale develop- ment and its influence on privacy protection behaviors and information verification behaviors. Telecommunications Policy, 50(2), Article 103117. https://doi.org/10.1016/j.te lpol.2025.103117 Yang, S., & Appleget, C. (2025). Besties with bots: GenAI as a lesson plan evaluator for preservice teachers. Journal of Digital Learning in Teacher Education, 41(2–3), 101–115. https://doi.org/10.1080/21532974.2025.2474932 Yim, I. H. Y., & Su, J. (2025). Artificial intelligence literacy education in primary schools: A review. International Journal of Technology and Design Education, 35, 2175– 2204. https://doi.org/10.1007/s10798-025-09979-w Yu, X., & Pian, Y. (2025). Towards actionable GenAI- classroom integration: A social role framework for teacher’s practice. In A. I. Cristea, E. Walker, Y. Lu, O. C. Santos, & S. Isotani (Eds.), Artificial intelligence in education. Posters and late breaking results, work- shops and tutorials, industry and innovation tracks, prac- titioners, doctoral consortium, blue sky, and wideAIED. AIED 2025 (Vol. 2590, p. 158–167). Springer. https: //doi.org/10.1007/978-3-031-99261-2_15 Zhang, Y., Lai, C., & Gu, M. M. Y. (2025). Becoming a teacher in the era of AI: A multiple-case study of pre- service teachers’ investment in AI-facilitated learning-to- teach practices. System, 133, Article 103746. https://doi. org/10.1016/j.system.2025.103746 Zhou, C., Ahmadi, S., Li, L., & Hossain, S. (in press). Con- ceptualizing generative AI literacy in teacher education: A systematic review. Author Contributions Shahin Hossain: Conceptualization, Methodology, Investi- gation and Validation, Writing – original draft (Section 4), Writing – review & editing, Proofreading. Sima Ahmadi: In- vestigation (literature review), Writing – original draft (Section 3, with contributions to other sections), Writing – review & editing. Leqi Li: Conceptualization, Methodology, Investiga- tion (literature review), Writing – original draft (contributions to Section 3), Writing – review & editing. Idowu David Awoyemi: Conceptualization, Methodology, Investigation and Validation, Writing – original draft (Sections 1, 6, and 7), Writing – review & editing. Wei Huang: Conceptualization, Writing – original draft (Section 5), Reference curation and in- text citation verification, Writing – review & editing. Chenxi Zhou: Investigation (literature review), Writing – original draft (contributions to Section 3 and other sections), Writing – review & editing, Reference and in-text citation verifica- tion. Jujia Li: Conceptualization, Writing – original draft (Section 5), Writing – review & editing. Samaa Haniya: Writ- ing – review & editing, Validation, Expert Feedback. Shapla Khanam: Writing – review & editing, Validation. Tasbirun Mashreka Subaha: Reference curation and in-text citation verification. All authors contributed to critical revision of the manuscript and read and approved the final version. All authors agree to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. Declaration of generative AI and AI-assisted technologies in the writing process During the preparation of this work, the authors used Claude (Anthropic; Fable 5, Opus 4.8, and Sonnet 4.6) and ChatGPT (OpenAI; GPT-5.5) to improve language clarity and style; copyedit the manuscript; check and correct reference format- ting and bibliographic details; verify internal consistency of citations, terminology, and cross-references; format tables and prepare figures; and check compliance with journal formatting requirements. All conceptual framing, arguments, analyses, and interpretations are the authors' own. After using these tools, the authors reviewed, verified, and edited all content and took full responsibility for the content of the published article. 20