自我调节学习对于学业成功和终身学习至关重要。然而, 如何避免人工智能对学习者主体性的过度替代, 并在人机交互中实现自我调节学习能力的培养, 成为当前研究的核心挑战。本文提出了以主体性为核心的三种人机协同模式, 系统梳理人工智能在自我调节学习中的研究现状。分析发现, 支架引导调节虽然能显著提升效率, 但也可能因过度替代引发依赖与元认知懒惰；共同决策调节能显化元认知过程, 却面临责任模糊和负荷加重风险, 工具辅助调节利于巩固自主感, 但可能放大能力鸿沟和技术不平等。未来的研究可考虑探讨人工智能与学习者主体性的关系和合理运用方式, 结合多模态和脑影像技术揭示人工智能赋能自我调节学习背后的心理过程，并通过纵向设计探索人工智能支持下自我调节学习能力的发展与变化。

Self-Regulated Learning (SRL), the capacity of learners to actively plan, monitor, and adapt their learning behaviors to achieve academic goals, is essential for academic success and the development of lifelong learning competencies. Yet many learners struggle to develop SRL. Traditional interventions are limited by scarce resources and delayed feedback, preventing them from providing sustained and individualized support. Artificial Intelligence (AI) offers potential solutions, but its specific positive and negative effects on students’ SRL remain underexplored. This paper reviews SRL theories and components, proposes three human-AI collaborative models centered on learner agency, analyzes AI’s mechanisms and implementation pathways across SRL stages, and discusses challenges and future directions.

Building on Bandura’s foundational work, researchers have developed influential models such as Zimmerman’s cyclical model, Winne and Hadwin’s model, and Pintrich’s four-phase model. Despite different emphases, these models view SRL as a dynamic, cyclical process involving learning goals and strategies, with cognition, metacognition, and motivation as its core components. Drawing on perspectives such as hybrid human-machine regulation and the “collective brain,” this paper examines how these components are externalized and reconfigured in AI-enabled environments. In such contexts, some metacognitive monitoring and cognitive strategy selection are transferred to AI as an external regulatory system. Learners must still selectively interpret and integrate AI feedback according to their goals and values, transforming SRL from an internal into a multi-source regulatory process. The central challenge is balancing human and machine regulation so that AI enhances rather than undermines SRL.

The paper proposes an agency-centered human-AI collaborative SRL framework with three modes: (1) AI-dominated scaffolded regulation, in which AI largely handles task analysis, path planning, and process monitoring while learners follow recommended pathways; (2) Human-AI co-regulated decision-making, in which learners and AI negotiate goals, strategies, and evaluations through dialogue while learners retain primary decision authority; and (3) Self-directed tool-supported regulation, in which learners with clear goals use AI on demand as an information and problem-solving resource while maintaining autonomy over planning, monitoring, and reflection. These modes intersect with the phases of forethought, performance, and reflection to form a “Collaborative Mode × SRL Phase” matrix showing the human-AI division of labor across phases and its implications for SRL development.

The paper then reviews empirical studies on AI-enhanced SRL in higher education from the past five years. AI can support SRL through personalized learning pathways, real-time feedback, process-data visualization, and the externalization of metacognitive activities. AI-scaffolded regulation can reduce cognitive load and improve short-term performance, especially for learners with weaker SRL skills. However, prolonged outsourcing of planning, monitoring, and evaluation may encourage dependency, superficial processing, and metacognitive laziness, ultimately weakening learner agency. Human-AI co-regulation can strengthen metacognitive monitoring, motivation, and autonomy when roles and decision boundaries are clear and dialogue is well structured. Poor design, however, may blur responsibility, encourage direct-answer seeking, and overwhelm learners with frequent prompts, increasing emotional burden and weakening sustained motivation. Tool-supported regulation is most conducive to transferable SRL competencies when learners possess strong self-regulatory abilities and AI literacy, but it may also widen competency gaps and digital inequality.

Effective human-AI collaborative SRL depends on feedback that is descriptive rather than directive, scaffolding that fades rather than persists, interaction that is bidirectional rather than unidirectional, and support that is available on demand rather than imposed. The core principle is that AI should strengthen, not replace, learners’ monitoring and regulatory capacities. This principle supports the integration of adaptive mode switching, scaffolding fading, transparency, and controllability into AI-supported SRL interventions and tools. Nevertheless, the three modes are theoretical ideal types and may not fully capture dynamic changes caused by tasks and cognitive load in authentic contexts. Because the framework is designed specifically to enhance university students’ SRL, it should not be automatically generalized to efficiency-oriented or other human-computer collaboration contexts.

Finally, the paper identifies three research directions. First, future studies should validate the three collaborative models and their transition mechanisms, test the framework’s generalizability, develop context-sensitive theoretical frameworks and agency measures, and examine interactions between cognition and affect. Second, multimodal learning analytics should be combined with neuroimaging techniques, including eye-tracking, fNIRS, and fMRI, to investigate the cognitive and neural mechanisms of human-AI co-regulation and detect early signs of overreliance. Third, longitudinal studies should examine the development, transfer, and possible regression of AI-supported SRL competencies, while expanding evaluation to include process indicators such as reflective depth.