关键词: 生成式学习, 生成式人工智能, 动态协同模型, 具身认知

Abstract: Generative artificial intelligence (GenAI) can improve learners' immediate performance while also creating risks for independent thinking, deep understanding, and sustained self-regulation. This apparent contradiction raises a theoretical question: why can similar forms of AI support lead to different, and sometimes opposing, learning outcomes? This article proposes a dynamic synergy model of AI-supported generative learning to explain these divergent effects. Rather than treating motivation, emotion, cognition, metacognition, and creative activity as isolated predictors of learning, the model examines how these processes continuously influence one another during human-AI interaction. Its central contribution is to conceptualize generative learning as a changing system in which external AI support reshapes the relationships among psychological processes without replacing the learner as the agent of meaning construction.
The model integrates five psychological processes into three interrelated dimensions. The first is higher-order metacognitive regulation, through which learners monitor their understanding, evaluate information and feedback, and adjust strategies. The second is the coordinated functioning of cognitive, affective, and volitional processes. Cognition supports the selection, organization, and integration of information; emotion shapes the experience of challenge, uncertainty, and progress; and volition sustains goals, effort, and engagement. The third is the embodied generation of meaning through creative activity. Learners develop and reorganize ideas through exploration, task performance, and feedback from relevant environments. These dimensions are not sequential stages. Metacognitive judgments can change motivation and cognitive effort, new understanding can alter emotional responses, and creative action can generate experiences that initiate further cycles of reflection and learning.
A further innovation concerns the position of AI within this system. GenAI can simultaneously function as an external cognitive resource, an interactive partner, a participant in distributed cognitive work, and a source of support for creative exploration. Its feedback quality, prompting style, degree of answer completion and automation, and mode of interaction can influence how learners allocate effort and responsibility. The model therefore distinguishes the availability of AI-generated information from the learner's actual engagement in meaning construction. High-quality explanations or fluent outputs do not automatically produce understanding. Their educational value depends on whether learners critically examine, reorganize, and test the information rather than merely accept it.
On this basis, the model proposes two theoretically distinct but potentially reversible pathways: scaffolding facilitation and cognitive substitution. In the scaffolding facilitation pathway, AI reduces auxiliary demands and offers prompts or explanatory feedback while learners retain responsibility for essential processes such as reasoning, comparison, evaluation, and self-explanation. Metacognitive monitoring and task-appropriate experiential feedback help transform external assistance into independently usable knowledge and creative capability. In the cognitive substitution pathway, AI increasingly performs the very operations learners need to practice. Learners may obtain polished answers and complete tasks efficiently while engaging less in independent analysis, monitoring, and verification. Thus, improvements in immediate output quality may coexist with weaker knowledge retention, transfer, or autonomous performance. These pathways describe hypothesized configurations of learning processes, not fixed categories of learners or technologies.
The allocation of cognitive responsibility is proposed as a central mechanism through which the two pathways are maintained or transformed. Cognitive offloading is not inherently harmful: delegating routine searches or formatting can preserve resources for meaningful learning, whereas outsourcing essential interpretation and judgment may undermine it. Pathway transitions may occur when learners change how they question AI, verify its suggestions, or reclaim responsibility for key decisions. Such changes are expected to depend jointly on AI literacy, prior knowledge, self-regulatory capacity, support design, and task demands. Embodied experience is also important, but its form should match the task: practical action and environmental feedback are especially relevant where learning requires application, experimentation, or situated judgment.
The model yields five directions for empirical testing. First, the effects of AI feedback should vary with learners' active processing and metacognitive regulation. Second, different allocations of cognitive responsibility should produce different learning outcomes even at comparable levels of automation. Third, metacognitive monitoring should help explain the persistence of, and transitions between, facilitative and substitutive learning states. Fourth, creative exploration and task-appropriate experiential verification should influence whether AI-supported outputs become personally understood and transferable knowledge. Fifth, immediate task performance and longer-term independent learning should be assessed separately because they may diverge. These propositions can be examined through experimental manipulations of feedback and automation, process-oriented analyses of human-AI interaction, longitudinal designs, and assessments conducted after AI assistance is withdrawn.
Overall, the dynamic synergy model moves beyond asking whether GenAI is beneficial or harmful. It asks how external intelligent support reorganizes the learner's psychological system, under what conditions support becomes substitution, and how learners can preserve agency while benefiting from AI. By connecting metacognitive regulation, cognitive-affective-volitional coordination, embodied creative activity, and cognitive responsibility within one framework, the model offers a theoretically integrated and empirically testable account of the dual effects of AI-supported generative learning.

Key words: generative learning, generative artificial intelligence, dynamic synergy model, embodied cognition