AI and self-regulated learningConnecting support, learner readiness and independent performance

Research question

The research question is how AI support can contribute to learners’ ability to plan, monitor and evaluate their own work across learning tasks. A useful reading route connects evidence at three levels: patterns across interventions, the design of a learning environment, and the capabilities learners bring to AI use. Reading the studies together suggests treating the learning phase, the support offered and the outcome measured as separate design decisions. This makes the question more precise than asking whether an AI tool improves learning in general. The proposed connection between these levels is a synthesis developed in this guide.

Evidence and study design

Xu and colleagues’ meta-analysis provides the starting point. It synthesizes 35 empirical studies published from 2013 to 2025 within Zimmerman’s cyclical model and reports an overall effect of g = 0.507. The reported effects differ across forethought, task performance and reflection, and across measurement approaches. This contribution is useful when motivating a study that locates AI support within a particular learning phase or compares different kinds of outcome evidence. The synthesis also directs attention to whether improvements during supported activity persist as independent self-regulatory competence. This is an assessment question that can structure the next empirical study.

That question can be translated into a concrete study plan. A researcher could examine how learners formulate a goal before consulting an agent, revise a strategy during a task and explain the changes after completing it. Immediate task scores, process records and a later task completed independently would answer complementary questions. These are proposed design choices derived from the meta-analysis’s emphasis on phases and competence. For example, a successful revision during assisted work can be examined alongside the learner’s ability to explain the relevant strategy and apply it in a new problem. The design connects the availability of support with evidence of what the learner can subsequently do.

MetaClassroom adds a perspective on the organization of the learning environment. This quasi-experimental study compares an immersive programming environment designed around three-dimensional learning progression with traditional classrooms. It reports improvements in programming achievement and self-regulated learning, including metacognitive skills, persistence and help-seeking, alongside learning beliefs and motivation. Its contribution to this topic is the joint consideration of disciplinary learning and regulation within an instructional design. It can be cited when explaining why a programming intervention evaluates multiple learning dimensions or integrates teaching, learning and assessment. The intervention studied is MetaClassroom; its findings concern that implemented environment.

Learner readiness and agent roles

Learner readiness enters through Jing and colleagues’ ChatGPT programming study. Using a quasi-experimental design with Python graphing tasks, it connects effectiveness in ChatGPT-assisted problem solving with AI literacy, programming knowledge and understanding of ChatGPT. Its contribution is to identify learner characteristics that matter within a concrete task setting. Read alongside MetaClassroom, it suggests a research design that describes both the environment and the learner’s starting point. A proposed follow-up could examine whether the same regulatory prompt is useful to students with different programming foundations, with those foundations measured before the learning activity.

Liu and colleagues’ review of GenAI agents offers a vocabulary for describing the support itself. Its synthesis distinguishes cognitive-epistemic, self-regulatory, affective and human–AI interaction roles. These categories help a researcher state what an agent is intended to contribute during learning. In a proposed programming activity, an agent could ask a learner to explain a debugging plan, review progress or evaluate a completed solution. Assigning such functions explicitly would make the intervention easier to describe and compare. The review supports the role classification; the example activities here are design proposals developed from that classification.

Research directions

These papers support a connected reading strategy. Use the meta-analysis to establish why phases and measurement matter; use MetaClassroom to examine the integration of disciplinary and regulatory outcomes; use the ChatGPT study to motivate attention to learner readiness; and use the agent review to specify educational roles. Together, they motivate a question that can guide a focused project: which form of agent support, introduced at which learning phase, helps learners with a specified starting level develop strategies they can use independently? A study organized around this question can define its intervention, comparison and assessment before selecting the technology. The relationship proposed here is a research agenda, with each paper supplying a distinct part of its rationale.