Young People, Learning, and Generative AI: A Rapid Literature Review and Implications for PreK-12 Education

Synthesis: Arthars, Yang, Hill, Liu, and Markauskaite (2026) — a rapid literature review from the University of Sydney synthesizing 271 empirical papers (2022–June 2026, Scopus + Web of Science) on how generative AI (GenAI) shapes learning for young people in PreK-12 settings. Framed by the How People Learn II ecology (learners, contexts, cultures) and an ecological learning-sciences view, the review finds no single "GenAI effect": affective gains are common but weak indicators of learning; the most consistent evidence is for improved immediate performance and product quality, while evidence on durable learning, transfer, and sustained self-AI Regulation in Education is more uneven. Outcomes depend on five entangled conditions — the learner, the tool, the task, social arrangements, and cultural/institutional context. The report's central contribution is a framework distinguishing students' surrender, offloading, or agency across cognitive, metacognitive, and affective dimensions, and four pedagogical functions of GenAI (learn from, with, about, or by shaping it).

Purpose, framing, and method

The review is aimed at educators, school leaders, and policymakers facing decisions that cannot wait for a mature evidence base. It takes a learning-sciences, ecological perspective: learning emerges from interactions among learners, tools, environments, and cultural norms, so GenAI's educational value cannot be assessed in isolation but only in relation to how it is configured.

• Method: rapid literature review (Garritty method), Scopus + Web of Science, 2022 to 30 June 2026, English-language peer-reviewed journal articles and conference proceedings; 271 papers included after screening; data coded across the learner/context/culture dimensions; retraction and publication-integrity checks run before finalization.
• Framing: the How People Learn II (HPL II) structure — the characteristics/capacities of learners, the contexts of learning, and the cultural dimensions that shape what learning means and who participates.

Key findings on outcomes

The evidence does not support a simple beneficial-or-harmful conclusion. Reported outcomes cluster into:

• Affective gains are common but weak. GenAI use is often associated with increased interest, engagement, Motivation, domain-specific Self-Efficacy, and reduced learning-related anxiety — across ages, subjects, and national contexts. But these are weak indicators of learning: confidence, ease, and perceived usefulness do not necessarily translate into durable understanding.
• Immediate performance improves most consistently. Gains are strongest in immediate task performance, product quality, and knowledge assessed shortly after GenAI use (writing grammar/structure, math practice, programming performance). However, stronger performance during GenAI-supported activity does not mean students developed retainable, transferable capability.
• Higher-order reasoning is promising but uneven. Studies associate GenAI-supported activities with argumentation, perspective-taking, critical reflection, evaluation, causal explanation, computational thinking, and creative Problem Solving — but these outcomes are less consistently demonstrated and more likely when activities require students to explain, justify, compare, evaluate, reflect, or revise.
• Metacognitive and self-regulated learning outcomes are mixed. GenAI can prompt monitoring/reflection/checking/revision during a supported activity, but benefits often do not persist when support is removed. Some uses encourage answer-seeking, dependence, weak verification, reduced self-monitoring, and a tendency to overestimate learning.
• The miscalibration gap. GenAI can increase perceived learning even when durable learning is not demonstrated (e.g., note-taking outperformed GenAI alone on retention, yet students preferred GenAI and saw it as more helpful) — consistent with the fluency bias where learning that feels fluent is often shallow.

Five entangled conditions shaping outcomes

1. Learner characteristics/orientations — prior knowledge and proficiency, self-AI Regulation in Education, Self-Efficacy, Help-Seeking and trust, and how students position GenAI in the activity. Higher-performing students use GenAI more strategically; lower-performing students outsource core work. The review highlights metacognitive inequity: weaker metacognitive students are more susceptible to detrimental offloading and less able to recognize it.
2. GenAI tool design — open vs. constrained, scaffold- vs. answer-oriented, domain-specific vs. general-purpose, personas/roles/interfaces. The clearest demonstration is Bastani et al. (2025): the same GPT-4 configured as an unrestricted GPT Base versus a guardrailed GPT tutor — students with GPT Base used it as a crutch and performed worse than controls when it was removed, while GPT Tutor students performed like controls. Productive friction built into tools (withholding answers, prompting explanation, Socratic questioning) supports learning.
3. The learning task — what students are asked to produce and how. Tasks are more supportive when GenAI reduces peripheral barriers (syntax, retrieval) rather than replacing the learning goal itself. Pedagogical functions matter: learning from GenAI (tutor), with GenAI (cognitive tool), about GenAI (AI literacy), or by shaping GenAI (teachable-agent).
4. Social arrangements — who participates and how. Hybrid teacher–GenAI feedback arrangements often outperform either alone; teacher orchestration, mediation, and Scaffolding are central to critical engagement. Peer collaboration evidence is more uneven.
5. Cultural and institutional contexts — resources, infrastructure, prior technology experience, cultural identity, and authority norms shape GenAI use and trust. Evidence here remains limited, so cross-context generalization needs care.

The surrender–offloading–agency continuum

Building on Shaw and Nave's (2026) Cognitive Surrender, the review's central practical framework distinguishes how students relate to GenAI across cognitive, metacognitive, and affective dimensions:

• Surrender — responsibility for learning-relevant work shifts to GenAI, often without deliberate awareness (accepting answers unchecked; letting GenAI decide what/how/when to learn; mistaking a fluent interaction for understanding).
• Offloading — the learner intentionally delegates selected work (summarizing after an initial attempt, using GenAI for a study plan). This can support performance but only becomes learning if the student checks, elaborates, and connects output to their own understanding.
• Agency — the learner retains responsibility for effort, judgment, and learning (asking GenAI to challenge an argument, generating practice scenarios, deliberately choosing not to use it when retrieval/effort is the point).

These are not fixed properties of a practice but of how a practice is enacted. This connects directly to the knowledge base's treatment of adaptive vs. maladaptive offloading and Reducing AI Misuse.

What this means for practice

• Instructors. Name the pedagogical function before choosing the tool — learning from, with, about, or by shaping GenAI — and judge the activity by what students remain responsible for: reading, reasoning, explaining, checking, revising, judging, and monitoring.
• Instructors. Keep productive friction in the task by having tools withhold answers and prompt explanation, and evaluate with evidence beyond the immediate product (process records, oral explanations, delayed or transfer tasks, unaided performance), because gains are concentrated in immediate performance and learner-reported confidence.
• Instructors. Teach regulation of GenAI explicitly and stage it to age and prior knowledge — when to use it, how to monitor its effect on thinking, when independent effort is the point — since metacognitive inequity leaves weaker self-regulators most exposed to detrimental offloading. GenAI literacy is necessary but not sufficient: students also need the foundational domain knowledge to question, critique and judge what a tool produces.
• Learners. Delegate after an initial attempt, then verify, elaborate, and connect the output to your own understanding; treat a fluent interaction as a prompt to check comprehension rather than proof that it landed, because perceived learning rises even when retention does not.
• Administrators. Replace adoption-or-ban with developmentally responsive guidance and fund the training and AI Governance hybrid human–GenAI arrangements demand: teacher–GenAI combinations often outperform either alone but require more teacher expertise, and access, privacy, safety, and procurement must be handled together.

Limitations

• Rapid review, not a systematic review or meta-analysis: 271 papers from Scopus and Web of Science (2022 to 30 June 2026), restricted to English-language peer-reviewed journal articles and conference proceedings plus selected Gray literature, with no pooled effect sizes, so the review cannot say how large any effect is.
• The authors describe the peer-reviewed base as "emergent and uneven in methodological rigour" and analyze it as a narrative, coded synthesis across the learner, context, and culture dimensions of How People Learn II rather than through a meta-analytic protocol.
• Outcome evidence is asymmetric: affective gains and immediate performance dominate the corpus while durable learning, transfer, and sustained self-regulation are inconsistently demonstrated, and the review notes that cultural and institutional evidence remains limited, so cross-context generalization needs care.
• Tool-design claims lean heavily on a single comparison (Bastani et al. 2025: unrestricted GPT Base versus a guardrailed GPT tutor) and on studies from a four-year window, so the design lessons are only as strong as that small set of trials.

Citation

Arthars, N., Yang, H., Hill, M., Liu, D., & Markauskaite, L. (2026). Young people, learning, and generative AI: A rapid literature review and implications for PreK-12 education. The University of Sydney.