The starting point for the framework is a diagnosis of what is missing from most technology-enhanced language learning environments. Reviews of artificial intelligence in language education document a striking expansion of capability, from automated assessment and tutoring to conversational practice and on-demand content generation. Generative AI in particular has added instant interaction and individualized support. But the same reviews repeatedly warn that effective and ethical use depends on digital competence, pedagogical framing, and human judgment. In many existing models, the AI tool is the most visible element while the actual learning mechanism is supplied by the task, such as progressive questioning, peer interaction, or repeated application. Without an explicit design, a large language model may generate polished prose while leaving the educational machinery vague or absent.

Zhang and Kim’s response is to specify exactly which component performs which function. The D-DSNM weaves together three traditions. Self-regulated learning theory, rooted in Zimmerman’s social cognitive model, describes a cyclical process in which learners engage in forethought by interpreting tasks and setting goals, regulate their performance by monitoring attention and strategy use, and reflect on outcomes to adapt future action. The Dual-Situated Learning Model, originally developed in science education to provoke conceptual change, contributes a logic of situated discrepancy: learners confront situations that expose a gap between their prior understanding and a target conception, then reconstruct and apply the revised understanding. Generative AI, in this scheme, is deliberately demoted to a supporting role as a generator and adapter of bilingual narrative material and reflective prompts.

The framework organizes learning around two interconnected situations. In the Narrative Exploration Situation, students encounter unfolding Chinese-English scenarios containing genuine communicative or cultural tension, for example a direct request that clashes with expectations about relational sensitivity, or an apparently equivalent expression that carries different pragmatic force. The point is not to teach a single correct cultural rule but to make assumptions visible and create a consequential reason to compare interpretations. In the Metacognitive Reflection Situation, the pace slows and attention turns inward: what assumption guided the initial response, which linguistic cue was overlooked, how might alternative wording change the interpersonal effect, and what strategy should be tried next? Insights recorded during reflection then feed into a new narrative episode, creating an iterative loop rather than a one-time sequence.

From this architecture, the authors derive three proposed learning mechanisms. The first runs from cognitive conflict to metacognitive awareness: a narrative event that does not fit the learner’s initial interpretation activates the forethought phase, prompting the learner to articulate a prediction, identify uncertainty, and set a goal for resolving the dilemma. The second involves adaptive scaffolding and strategy activation, in which graduated prompts move from noticing the main misunderstanding to comparing linguistic evidence, inferring a cultural perspective, testing a reformulation, and explaining why a strategy might work, with support gradually faded as learners demonstrate greater control. The third mechanism, cyclical reinforcement and strategy internalization, connects self-reflection to subsequent performance: a strategy identified during reflection is tested in a new narrative branch with altered linguistic and cultural conditions, helping learners distinguish a flexible principle from a memorized response.

These mechanisms are translated into five concrete design directives, each accompanied by sample prompts and classroom scenarios. Designers are told to calibrate culturally consequential dilemmas, adjusting linguistic and cultural complexity to learners’ readiness; one sample prompt asks the AI to generate a Chinese-English university negotiation in which preserving face conflicts with a direct request for clarification, offering two plausible interpretations rather than one correct cultural explanation. Scaffolds should be sequenced and faded, moving from noticing to interpretation to strategy selection to justification. Reflection should be externalized through logs, decision maps, or annotated dialogues that make strategy use visible. The interface should build an explicit return loop, allowing learners to pause a narrative, reflect, and re-enter a modified episode that demands application of the revised strategy. Finally, narrative branching must preserve learner choice, offering meaningful continuations linked to learner goals rather than silently optimizing a pathway.

A crucial feature of the framework is its insistence on the replaceability of the AI platform. The authors consulted the DeepSeek-V3 technical report only to describe baseline technical characteristics and benchmark performance, explicitly refusing to treat it as evidence of educational effectiveness. Benchmarks may support the feasibility of generating text for prototype activities, but they cannot establish that a system promotes self-regulation, metacognition, or bilingual competence. If another sufficiently capable language model can generate and revise suitable materials under the same pedagogical constraints, the theoretical sequence remains intact. This platform-independent logic distinguishes the framework from claims that derive educational innovation primarily from the novelty of a particular tool, and it positions AI as a configurable resource within a human-governed learning design.

The authors are equally candid about the risks and the limits of their evidence. The study is a critical literature synthesis and theory-driven conceptual analysis, not an empirical test; no primary data were collected, and every claim about the model is framed as a proposition requiring validation. The transfer of the Dual-Situated Learning Model from science conceptual change to bilingual pragmatic learning is a theoretical extension that may not operate as proposed. Generative outputs can contain hallucinations, cultural stereotyping, bias, and privacy risks, and experimental evidence suggests that generative AI assistance may improve immediate task performance without equivalent knowledge gain, sometimes fostering metacognitive dependence when learners are insufficiently guided. Implementation, the authors argue, would require data-protection procedures, disclosure of AI use, human review of generated scenarios, options to contest feedback, and criteria for culturally responsible design.

The proposed bilingual competence itself is carefully delimited. Rather than implying balanced, native-like mastery of two languages, the framework defines four observable domains: bilingual pragmatic performance, including requests, repair, stance, and facework; cross-linguistic strategy use, including comparison, reformulation, and mediation; cultural perspective-taking, understood as recognizing multiple plausible interpretations rather than treating culture as a fixed rule; and metacognitive regulation of bilingual choices, spanning planning, monitoring, explaining, and revising a response. These domains set the outcome boundaries and future indicators against which any empirical test of the model would be judged, and they give instructors something more precise to assess than fluency alone.

What comes next is a staged research agenda. Usability and design-based studies would first examine whether instructors and learners understand the two situations and can work with the proposed prompt sequences. Small-scale process studies would then analyze reflection logs, dialogue revisions, pathway choices, and prompt-response traces to determine whether the three mechanisms are actually observable in practice. Only after that would quasi-experimental or mixed-method studies compare D-DSNM-informed instruction with alternative technology-enhanced designs, using measures aligned with the four domains of bilingual competence alongside self-regulation, delayed transfer, and learner agency. Cross-platform replication and culturally diverse settings would be necessary before broader conclusions could be drawn. The central claim of the framework is deliberately modest: generative AI can help instantiate and adapt instructional materials, but the anticipated educational value depends on how educators design conflict, scaffolding, reflection, learner choice, and repeated application. In a moment when institutions are racing to deploy chatbots across curricula, that insistence on pedagogy over platform may prove to be the study’s most consequential message.

Subject of Research: A conceptual design framework integrating self-regulated learning theory, the Dual-Situated Learning Model, and generative AI-facilitated bilingual narrative exploration in university education

Article Title: Design principles for supporting self-regulation and bilingual competence through AI-facilitated narrative exploration in university students

Article References: Zhang, W., & Kim, Y. (2026). Design principles for supporting self-regulation and bilingual competence through AI-facilitated narrative exploration in university students. Discover Education, 5(1), Article 999. https://doi.org/10.1007/s44217-026-02211-4

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02211-4

Keywords: self-regulated learning, bilingual competence, generative AI, DeepSeek, Dual-Situated Learning Model, narrative exploration, metacognition, instructional design, language education, higher education, Chinese-English learning, educational technology