Tran Thi Hoang Anh
Hanoi University of Civil Engineering, Hanoi, Vietnam
Abstract
The integration of generative artificial intelligence (AI) in language education has expanded opportunities for instructional support, yet its pedagogical role remains insufficiently theorised. Existing research has largely focused on discrete applications, with limited attention to how AI can be systematically incorporated into lesson design. This study proposes a conceptual framework for integrating generative AI into EFL instruction in higher education, drawing on a narrative synthesis of EFL pedagogy, Cognitive Load Theory, and sociocultural theory. The framework positions AI as an embedded component of instructional design and is structured across three interrelated dimensions: teacher cognition and pedagogical design, cognitive regulation during task execution, and AI-mediated learner interaction. In this model, generative AI redistributes cognitive load and mediates learning processes within task-based instruction. The framework contributes to current discussions by offering a more integrated account of AI in language teaching and provides a foundation for more systematic and pedagogically grounded implementation in university contexts.
Keywords: generative artificial intelligence, EFL lesson design, cognitive load theory, sociocultural theory, instructional design, AI-mediated learning.
1. Introduction
The integration of artificial intelligence (AI) into education has intensified in recent years, reshaping pedagogical practices across disciplines. In EFL contexts, generative AI tools are increasingly used to support material development, feedback provision, and learner interaction.
In Vietnam, these developments occur within a policy-driven system aligned with CEFR B1 outcomes and characterised by large class sizes, heterogeneous proficiency levels, and exam-oriented instruction, where lesson design becomes a critical locus of pedagogical decision-making.
Despite growing adoption, the pedagogical integration of generative AI remains under-theorised. Existing studies largely focus on discrete applications, with limited attention to how AI can be systematically embedded within lesson design. This gap is consequential, as lesson design involves coordinating objectives, tasks, and learner variables in complex instructional contexts.
This study addresses this limitation by proposing a conceptual framework that positions generative AI as a mediational component within instructional design, linking teacher cognition, cognitive regulation, and learner interaction.
2. Literature Review
2.1 Generative AI in Language Education
The emergence of generative artificial intelligence (AI), particularly large language models (LLMs), has introduced a qualitatively different form of technological mediation in language education. Unlike earlier tools, which primarily supported discrete functions such as grammar checking or content delivery, generative AI systems can produce contextually appropriate and linguistically coherent text in real time (Dwivedi et al., 2023; Kasneci et al., 2023), enabling a more active role in instructional processes.
Within EFL contexts, generative AI is most commonly used for writing support, feedback provision, and material generation. It assists learners in idea development, organisation, and linguistic formulation (Nazari et al., 2021; Wang et al., 2024), provides immediate feedback to support revision (Ranalli, 2018; Li et al., 2023), and enables teachers to rapidly produce and adapt instructional materials under time constraints (Holmes et al., 2022; Luckin et al., 2022).
Despite these developments, the literature remains predominantly application-oriented, with limited attention to how AI is pedagogically integrated within lesson design (Godwin-Jones, 2023). As a result, its use often appears fragmented rather than systematically embedded in instructional structures. More fundamentally, there is limited understanding of how AI-mediated practices interact with core dimensions of lesson design, including task structuring, cognitive regulation, and learner engagement. This highlights the need for a more integrated conceptualisation of AI within instructional design.
2.2 Theoretical Foundations of Lesson Design in EFL
Established models of EFL pedagogy provide important foundations for understanding instructional design, yet offer limited guidance on the integration of generative AI. The Technological Pedagogical Content Knowledge (TPACK) framework conceptualises effective teaching as the integration of technological, pedagogical, and content knowledge (Mishra & Koehler, 2006; Koehler et al., 2013), emphasising alignment between technology use and pedagogical intent (Tai, 2015; Tseng et al., 2022). This principle is particularly relevant for generative AI, whose flexibility may encourage use without clear pedagogical grounding.
Task-Based Language Teaching (TBLT) and Communicative Language Teaching (CLT) further highlight the importance of structuring instruction around meaningful activity and interaction. TBLT organises lessons into sequences of task preparation, performance, and reflection (Ellis, 2003; Willis & Willis, 2007; Long, 2015), while CLT emphasises communicative use of language (Richards, 2006). Together, these approaches underscore that effective lesson design depends on the coordinated organisation of tasks, resources, and interaction.
However, these frameworks do not fully account for the affordances of generative AI, particularly its capacity for real-time content generation and adaptive interaction. While foundational, they require further development to support systematic integration of AI within lesson design.
2.3 Cognitive Perspectives on Learning and Instruction
Cognitive Load Theory (CLT) provides a Cognitive Load Theory (CLT) explains how instructional design shapes learning by regulating cognitive resources within the constraints of working memory (Sweller, 1988; Sweller et al., 2019). It distinguishes between intrinsic load (task complexity), extraneous load (ineffective design), and germane load (schema construction), with effective instruction aiming to optimise their distribution (Paas & van Merriënboer, 2020).
In EFL contexts, managing cognitive load is essential, as learners must process linguistic form, meaning, and use simultaneously. Practices such as task sequencing and scaffolding help regulate processing demands (Mayer, 2014), aligning with pedagogical approaches such as TBLT.
Generative AI alters this distribution by providing real-time input, feedback, and linguistic support. While it may reduce extraneous load and assist in managing intrinsic complexity, it also introduces new demands related to evaluating and integrating AI-generated output. Thus, rather than reducing effort, AI redistributes cognitive processing across different types of load.
Despite this potential, the implications of AI-mediated interaction for cognitive load remain under-theorised. Existing CLT-informed models assume relatively stable instructional conditions and offer limited guidance for dynamic, AI-supported environments, indicating the need for a more interaction-sensitive extension of CLT.
2.4 Sociocultural Perspectives on Language Learning
Sociocultural theory conceptualises learning as socially mediated activity, where development occurs through interaction and the use of mediational tools (Vygotsky, 1978; Lantolf & Thorne, 2006). Central to this perspective is the Zone of Proximal Development (ZPD), within which learners can perform beyond their current level with appropriate support.
Scaffolding enables this process through contingent assistance, typically realised in EFL contexts through interaction, feedback, and task design (Wood et al., 1976; Hammond & Gibbons, 2005).
Generative AI introduces a new form of mediation by providing on-demand input, feedback, and prompts. However, unlike human scaffolding, AI support is learner-initiated, continuously available, and not inherently contingent on pedagogical judgement. This raises questions about how mediation is regulated and how learning is shaped when support is partially transferred to non-human agents (Lantolf et al., 2015; van Lier, 2004).
Although sociocultural theory offers a well-established account of mediated learning, its application to AI-integrated instruction remains limited, particularly in explaining how AI operates within the ZPD or reshapes scaffolding processes.
2.5 Synthesis and Research Gap
While pedagogical, cognitive, and sociocultural perspectives address key aspects of lesson design, they remain insufficiently integrated in explaining how generative AI operates across these dimensions. This gap motivates the development of an integrative framework in which AI is embedded within instructional design.
3. Methodology
3.1 Research Design
This study adopts a conceptual research design oriented towards theory development. A narrative literature review is employed to synthesise and reinterpret relevant theoretical perspectives (Baumeister & Leary, 1997; Snyder, 2019), an approach suited to emerging areas such as generative AI in education where theoretical integration remains limited.
3.2 Analytical Approach
The analysis draws on a purposive selection of theoretically grounded works in language pedagogy, cognitive learning theory, and sociocultural perspectives, prioritising conceptual relevance over exhaustive coverage. An abductive analytic strategy is adopted to identify relationships among key constructs (Timmermans & Tavory, 2012), allowing iterative movement between existing theory and emerging insights. Through constant comparison, conceptual convergences and tensions are used to inform model development.
3.3 Framework Development Procedure
The framework is developed through iterative conceptual mapping, in which constructs are abstracted and reorganised according to their functional roles within lesson design. Following established approaches to conceptual framework construction (Jabareen, 2009), this process involves identifying core constructs, examining their relationships, and integrating them into a multi-layered model linking pedagogical design, cognitive regulation, and learner interaction. The resulting framework represents a theoretically informed synthesis rather than a direct aggregation of existing models.
4. Proposed Conceptual Framework
4.1 Overview of the Framework
To address the limitations identified in the preceding review, this study proposes the AI-Mediated Lesson Design Framework (AI-MLDF) for EFL instruction in higher education. The framework conceptualises generative AI as an embedded component of lesson design operating across pedagogical, cognitive, and interactional dimensions.
It is organised into three interrelated layers: teacher cognition and pedagogical design, cognitive regulation during instruction, and mediated learner interaction. These layers are treated as interdependent dimensions of AI-augmented lesson design rather than discrete components.
As illustrated in Figure 1, the framework follows a directional yet recursive logic. Lesson design begins with pedagogical structuring, is enacted through cognitively guided processes, and unfolds through learner interaction with tasks and AI-supported resources. Outcomes from this process inform subsequent design decisions, forming a cycle of ongoing refinement.
4.2 Layer 1: Teacher Cognition and Pedagogical Design
The first layer addresses how teachers make design decisions when incorporating generative AI into EFL lessons. Rather than treating TPACK as a static body of knowledge, it is operationalised as situated choices regarding task objectives, sequencing, and the use of AI support.
Lesson design involves determining where AI can meaningfully contribute, such as generating input, guiding task performance, or anticipating learner difficulties. These decisions are shaped by instructional goals and contextual constraints, including class size, learner variability, and time pressure.
Crucially, AI use is selective. Teachers allocate roles between teacher, learner, and AI to ensure that technology supports rather than displaces instructional intent. This layer establishes the structure of the lesson, shaping subsequent cognitive processing and interaction.
4.3 Layer 2: Cognitive Regulation through AI-Supported Instruction
The second layer specifies how generative AI reshapes the distribution of cognitive load during task execution. Drawing on Cognitive Load Theory (CLT), it focuses on how load is reduced, shifted, and intensified depending on how AI is embedded within lesson activities.
Generative AI can reduce extraneous load by clarifying instructions, modelling task expectations, and providing immediate examples. In EFL contexts, where instructional ambiguity can increase processing demands, such support enables learners to focus on task-relevant content. AI may also assist in managing intrinsic load by adapting input difficulty, for example through simplified texts or graded examples aligned with learner proficiency.
At the same time, AI redistributes rather than eliminates cognitive effort. By supporting lower-level processes such as idea generation, vocabulary retrieval, and sentence formulation, it offloads routine demands. However, this often leads to increased germane load, as learners must evaluate AI-generated output, select appropriate responses, and integrate feedback into their performance. Cognitive effort is thus reallocated towards higher-order processing.
The effects of this redistribution vary across task stages. In pre-task phases, AI may support schema activation and reduce initial processing demands. During task performance, it can provide on-demand scaffolding that stabilises cognitive load. In post-task phases, AI-mediated feedback may intensify germane load by prompting reflection and revision. Cognitive load is therefore dynamically regulated rather than fixed at the point of design.
Within the framework, AI functions as a cognitive regulator embedded in instructional processes. Its effectiveness depends on how task design anticipates and directs this redistribution, ensuring that reduced effort in one area is balanced by productive engagement in another.
4.4 Layer 3: AI-Mediated Learner Interaction
The third layer focuses on how learning is mediated through interaction during task performance, with generative AI functioning as an additional source of support alongside teachers and peers. In this layer, the emphasis shifts from task design and cognitive regulation to how learners engage with available resources to complete tasks and extend their performance.
Within AI-supported environments, mediation occurs through multiple channels. Learners may consult AI systems to clarify meaning, request reformulations, or generate language needed to complete a task. These interactions provide contingent support, allowing learners to access assistance at the point of need rather than relying solely on teacher input. As a result, the process of scaffolding becomes more immediate and continuously available throughout task performance.
However, AI-mediated support differs from traditional forms of scaffolding in both form and function. Unlike teacher-guided interaction, which is typically selective and socially negotiated, AI support is on-demand and learner-initiated. This shifts greater responsibility to learners in determining when and how to seek assistance, as well as how to interpret and apply the output provided. In this sense, mediation is not simply delivered to the learner but co-constructed through interaction with the system.
The effectiveness of this layer depends on how well AI use is aligned with task demands. For example, in communicative tasks, excessive reliance on AI-generated responses may reduce opportunities for meaningful language production, whereas strategically timed support can enable learners to sustain interaction and complete tasks beyond their current level. Thus, AI-mediated interaction must be calibrated to support, rather than replace, learner effort.
At this stage, AI shapes how learners sustain interaction and complete tasks. Its value depends on use: it may extend performance or displace it. Learner–AI interaction thus becomes the point at which instructional design is realised and evaluated.
4.5 Dynamic Feedback Loop and System Integration
The three layers operate as an integrated system in which lesson design is continuously informed by what occurs during task performance. As learners engage with tasks and AI-supported resources, their responses, patterns of support use, and points of difficulty reveal how instructional design functions in practice.
These observations provide the basis for adjustment. Teachers may modify task complexity, reposition AI support, or redesign interactional demands where cognitive load is either insufficient or excessive. The feedback loop is therefore necessary because AI-mediated instruction introduces variability that cannot be fully anticipated at the design stage.
Lesson design is thus not fixed in advance but stabilised over successive iterations. The framework captures AI-augmented teaching as a system in which design is progressively refined through use, rather than predetermined prior to instruction.
5. Conclusion
This study has proposed a conceptual framework for integrating generative artificial intelligence into EFL lesson design in higher education, positioning AI as an embedded pedagogical mediator within instructional processes. By drawing on pedagogical, cognitive, and sociocultural perspectives, the framework offers a theoretically grounded account of how AI shapes task structuring, redistributes cognitive demands, and mediates learner interaction in AI-supported environments.
The framework carries both theoretical and pedagogical implications. It advances current discussions by moving beyond application-oriented approaches towards a more integrated conceptualisation of AI in language education, while also highlighting the need for deliberate and context-sensitive integration in instructional practice, particularly in settings characterised by large classes and resource constraints.
Notwithstanding these contributions, the framework remains conceptual in nature and requires empirical validation. Future research should therefore examine its implementation in classroom contexts, with particular attention to its effects on cognitive load distribution, interactional processes, and learning outcomes.
References
[1] R. F. Baumeister and M. R. Leary, “Writing narrative literature reviews,” Rev. Gen. Psychol., vol. 1, no. 3, pp. 311–320, 1997. doi: 10.1037/1089-2680.1.3.311
[2] Y. K. Dwivedi et al., “So what if ChatGPT wrote it? Multidisciplinary perspectives on generative conversational AI,” Int. J. Inf. Manage., vol. 71, 102642, 2023. doi: 10.1016/j.ijinfomgt.2023.102642
[3] R. Ellis, Task-Based Language Learning and Teaching. Oxford, U.K.: Oxford Univ. Press, 2003.
[4] R. Godwin-Jones, “ChatGPT and the future of language learning,” Lang. Learn. Technol., vol. 27, no. 2, pp. 1–13, 2023.
[5] J. Hammond and P. Gibbons, “What is scaffolding?” in Teachers’ Voices 8, Sydney, Australia: NCELTR, 2005.
[6] W. Holmes, M. Bialik, and C. Fadel, Artificial Intelligence in Education. Boston, MA, USA: CCR, 2022.
[7] Y. Jabareen, “Building a conceptual framework,” Int. J. Qual. Methods, vol. 8, no. 4, pp. 49–62, 2009. doi: 10.1177/160940690900800406
[8] E. Kasneci et al., “ChatGPT for good? Opportunities and challenges,” Learn. Individ. Differ., vol. 103, 102274, 2023. doi: 10.1016/j.lindif.2023.102274
[9] M. J. Koehler, P. Mishra, and W. Cain, “What is TPACK?” J. Educ., vol. 193, no. 3, pp. 13–19, 2013. doi: 10.1177/002205741319300303
[10] J. P. Lantolf and S. L. Thorne, Sociocultural Theory and Second Language Development. Oxford, U.K.: Oxford Univ. Press, 2006.
[11] J. P. Lantolf, S. L. Thorne, and M. E. Poehner, “Sociocultural theory,” in Theories in Second Language Acquisition, 2nd ed. New York, NY, USA: Routledge, 2015.
[12] S. Li, L. Jiang, and X. Zhang, “AI-assisted feedback in L2 writing,” Comput. Assist. Lang. Learn., 2023. doi: 10.1080/09588221.2023.2167025
[13] M. H. Long, Second Language Acquisition and Task-Based Language Teaching. Oxford, U.K.: Wiley-Blackwell, 2015.
[14] R. Luckin et al., Intelligence Unleashed. London, U.K.: Pearson, 2022.
[15] R. E. Mayer, Cambridge Handbook of Multimedia Learning, 2nd ed. Cambridge, U.K.: Cambridge Univ. Press, 2014.
[16] P. Mishra and M. J. Koehler, “Technological pedagogical content knowledge,” Teachers College Record, vol. 108, no. 6, pp. 1017–1054, 2006.
[17] M. Nazari, M. S. Shabbir, and R. Setiawan, “AI tools in second language writing,” Educ. Inf. Technol., 2021. doi: 10.1007/s10639-021-10548-4
[18] F. Paas and J. J. G. van Merriënboer, “Cognitive load theory,” Curr. Dir. Psychol. Sci., vol. 29, no. 4, pp. 394–398, 2020. doi: 10.1177/0963721420922183
[19] J. Ranalli, “Automated written corrective feedback,” Comput. Assist. Lang. Learn., vol. 31, no. 7, pp. 653–674, 2018. doi: 10.1080/09588221.2018.1428236
[20] J. C. Richards, Communicative Language Teaching Today. Cambridge, U.K.: Cambridge Univ. Press, 2006.
[21] H. Snyder, “Literature review as a research methodology,” J. Bus. Res., vol. 104, pp. 333–339, 2019. doi: 10.1016/j.jbusres.2019.07.039
[22] J. Sweller, “Cognitive load during problem solving,” Cogn. Sci., vol. 12, no. 2, pp. 257–285, 1988. doi: 10.1207/s15516709cog1202_4
[23] J. Sweller, P. Ayres, and S. Kalyuga, Cognitive Load Theory. Cham, Switzerland: Springer, 2019.
[24] K. W. H. Tai, “Teacher knowledge for technology integration,” Educ. Technol. Soc., vol. 18, no. 4, pp. 234–246, 2015.
[25] S. Timmermans and I. Tavory, “Abductive analysis,” Sociol. Theory, vol. 30, no. 3, pp. 167–186, 2012. doi: 10.1177/0735275112457914
[26] J. J. Tseng, Y. S. Cheng, and H. C. Yeh, “TPACK in language classrooms,” Comput. Assist. Lang. Learn., 2022. doi: 10.1080/09588221.2022.2033788
[27] L. van Lier, The Ecology and Semiotics of Language Learning. Dordrecht, Netherlands: Springer, 2004.
[28] L. S. Vygotsky, Mind in Society. Cambridge, MA, USA: Harvard Univ. Press, 1978.
[29] Y. Wang, L. J. Zhang, and S. Li, “Generative AI in L2 writing,” System, 2024. doi: 10.1016/j.system.2023.102993
[30] D. Willis and J. Willis, Doing Task-Based Teaching. Oxford, U.K.: Oxford Univ. Press, 2007.
[31] D. Wood, J. S. Bruner, and G. Ross, “The role of tutoring,” J. Child Psychol. Psychiatry, vol. 17, no. 2, pp. 89–100, 1976. doi: 10.1111/j.1469-7610.1976.tb00381.x