A Systematic Review of Generative AI in K–12: Mapping Goals, Activities, Roles, and Outcomes via the 3P Model

Abstract

Generative AI is reshaping k–12 learning as a multi-agent system in which goals, activities, and roles co-evolve across formal and informal environments. Following PRISMA and appraising quality with MMAT, we synthesize 84 peer-reviewed empirical studies (2020–2025) involving learners aged 3–18. Using Biggs’s 3P model as a systems lens and embedding CIMO logic, we code learning objectives, activity designs, AI role paradigms, and outcomes. Seven recurring objectives emerge (language/literacy; STEM; creativity; socioemotional skills; feedback literacy and self-regulation; motivation; AI literacy). Five dominant activity patterns are identified: dialogic tutoring and formative feedback, generative iterative co-creation, project-based problem-solving, simulation/game-based learning, and assessment support. Across studies, AI roles shift from AI-directed to AI-supported/empowered, re-allocating agency among students, teachers, and caregivers via feedback loops. Reported outcomes span three categories—epistemic, practice, and affective/identity—with opportunities of deeper knowledge, improved practice, and stronger engagement, and risks of hallucinations, reduced originality, over-reliance, motivational loss, and ethical concerns. We propose a goal–activity–role alignment heuristic for instructional design, plus safeguards around teacher professional development, feedback literacy, and ethics. We call for longitudinal and cross-cultural research to evaluate the impacts of GenAI in k–12.

1. Introduction

Over the past three years, generative AI (GenAI) has driven a system-level transformation of k–12 learning ecosystems while surfacing persistent challenges. Nearly half of 13–18-year-olds report using GenAI to write, draw, or produce videos for schoolwork or leisure [1]. Powered by large-scale deep learning models, these systems can generate text, images, audio, code, and other modalities from natural-language prompts [2]. Their functions are now embedded in widely used learning platforms and social applications, offering young learners personalized support and multimodal creative tools. Students can edit text in real time, pose questions, and receive instant feedback [3], which broadens forms of expression, strengthens creativity [4,5], and supports higher-order and critical thinking [6,7]. Teachers and parents are likewise introducing conversational AI into classrooms and homes—to generate teaching materials, orchestrate chained questioning, co-author stories, and automate grading [8,9]. Compared with traditional AI, GenAI enables educators, caregivers, and students in the education system to interact with content in new ways that better respond to heterogeneous learner needs and preferences [10]. However, these gains come with significant ethical and epistemic risks. Hallucinations, algorithmic bias, over-reliance, and privacy concerns raise ethical and epistemic issues that can shape children’s cognitive and socioemotional development [11,12]. Children often overestimate GenAI’s capabilities [13,14]; when the system produces biased or incorrect information, they may accept and reinforce it [15]. Excessive dependence can discourage deep thinking or even facilitate academic misconduct, undermining achievement, perceived learning, and higher-order cognition [16,17]. Because k–12 learners are in critical periods of cognitive, emotional, and social growth, AI tools designed for them must be developmentally appropriate [18]. Primary and lower-secondary classrooms feature diverse goals, tasks, and trajectories. More than rote recall, effective learning entails abstract reasoning and metacognitive monitoring [19]. Yet systematic frameworks that verify how well GenAI addresses these varied needs are still scarce [8]. At the same time, students’ mental models of AI shift with age, from tool to partner/collaborator, which affects how they adopt and internalize AI feedback [20]. Teachers are likewise moving from serving as the sole source of knowledge to acting as facilitators and co-learners [21]. In informal settings, parents use AI to tailor content and provide immediate feedback, satisfying children’s momentary interests, strengthening parent–child interaction, and promoting social participation [22]. As those who best understand a child’s cognitive level, caregivers can also supply AI systems with targeted contextual information [23]. Clarifying how young people interact with GenAI, what learning outcomes result, and how teachers and caregivers should collaborate within this ecosystem has therefore become a pressing agenda for both schools and families. As shown in Figure 1, we model K–12 GenAI use as a sociotechnical activity system. GenAI generates and adapts; adults curate tasks, norms, and context; and learners prompt, evaluate, and appropriate outputs. These feedback loops co-regulate the activity and anchor our analytic lens for the review [24,25].

The field lacks a theoretically led classroom-operational account that links learning goals, pedagogical activities, and human–AI role configurations to measurable outcomes for 3–18-year-olds. To address this gap, we synthesize 84 K–12 studies using an integrated 3P and CIMO lens to link goals, activities, AI roles, and outcomes. We contribute a theory-led goal–activity–role–outcome map, a clarified shift toward co-regulated human–AI ensembles, and practice-oriented guidance.

2. Literature Review

Empirically, although GenAI is rapidly gaining relevance for children and adolescents, the AIEd corpus remains skewed toward higher education; rigorous K–12 empirical studies remain sparse in both number and depth [27,28,29]. AI research in k–12 continues to privilege traditional AI (e.g., ITS) [30], emphasizing tool properties [31,32] or AI literacy [31,33,34] instead of tracing how AI is pedagogically enacted to enhance learning. Most of that work also predates the public release of powerful large language models, and therefore cannot capture GenAI’s distinctive affordances and risks. A separate stream of reviews surveys GenAI’s administrative, technical, and outcome-focused uses in teaching [35,36,37,38]. Taken together, these strands acknowledge AI’s importance for K–12, yet they largely remain at the level of “assistive analytics and recommendation.” By contrast, GenAI’s capacity to generate text, images, code, and other multimodal content [39,40,41] opens genuinely new space for personalization, learner–AI co-creation, and immersive assessment—while simultaneously surfacing fresh questions about role allocation, ethical governance, and how to evaluate impact at scale [42]. This leaves open a central question: How, concretely, does GenAI reconfigure classroom and home learning processes for 3–18-year-olds?

Preliminary reviews have begun to address this gap but still leave blind spots. Marzano (2025) [30] synthesized 197 empirical studies and concluded that GenAI can foster personalization, motivation, and assessment innovation; however, cross-disciplinary evidence and classroom-level implementations remain rare, and comprehensive teacher training and policy frameworks are underdeveloped [43]. Other analyses from teachers’ or leaders’ vantage points offer policy guidance but little operational direction for classrooms and families [44]. Zhang and Tur (2024) [28] cataloged promises and risks and called for co-designed, multi-stakeholder research; Miao et al. (2024) [45] highlighted content generation and dialogic collaboration. Nevertheless, these accounts stay largely at a macro level and fall short of offering fine-grained analyses of classroom interactions and measured learning outcomes.

Social learning theories remind us that learning unfolds in classrooms, homes, and online communities [32,46], yet we still lack a systematic picture of how children, parents, and teachers interact with AI in these intertwined settings. Moreover, from a learning sciences standpoint, learning is a complex, situated, generative, and embodied–agentive process [24] that is distributed across people, tools, and communities [24]; accordingly, we need to move beyond the knowledge-acquisition metaphor [47] toward explanatory frameworks of participation, collaboration, and distributed cognition [48].

In role theorization, AIEd has long positioned AI as a tool or assistant, foregrounding automation and data-driven decision-making and casting students as passive users [49]. Ouyang and Jiao (2021) [50] proposed three paradigms (AI-directed/AI-supported/AI-empowered), Xu and Ouyang (2022) [51] introduced a complexity-theoretic perspective; yet both are mainly grounded in traditional AI and, in essence, delineate a spectrum of learner control. Distributed cognition holds that cognitive processes are distributed across mind, body, and artifacts, implying that the unit of analysis is no longer the isolated learner but the integrated human–machine system [52]. With GenAI, the unit of analysis shifts from the isolated learner to human–AI ensembles, calling for role taxonomies that move beyond tool/teammate labels to account for co-regulation and distributed cognition [53]. Although recent reviews add labels such as Generator, Facilitator, Collaborator, Enhancer [54]; they underspecify learning goals beyond creative production and underattend key risks.

Likewise, the traditional knowledge–skills–attitudes/values outcomes framework [55] reinforces an acquisition metaphor of learning. This perspective, which views learning as the accumulation of discrete information, is insufficient for capturing the complex, generative processes facilitated by GenAI. Constructivist, situated, and sociocultural perspectives define learning as active meaning-making, participation in communities of practice, and identity development [24,25,56]. Consequently, to understand the impact of GenAI, we need a framework that can capture changes in learners’ ways of knowing, being, and doing.

Against this backdrop, this study addresses these empirical and theoretical gaps. We move beyond prior tool-centric syntheses or macro-level commentaries by providing a comprehensive and theoretically grounded account of how GenAI is being used in K–12 settings. By integrating the 3P model with the CIMO framework and operationalizing them with learning theories appropriate to the GenAI era, we build a systematic analytic lens to examine the interplay among learning goals, pedagogical activities and roles, and learning outcomes. This approach allows us to not only map the current landscape but also to advance the theoretical conversation, offering a practice-ready map that can inform curriculum design, teacher professional learning, and future research.

3. Materials and Methods

We conducted a systematic review guided by PRISMA [57] to identify, select, and synthesize empirical studies reviewed by peers on the use of GenAI in the education of children and adolescents (ages 3–18) in formal and informal settings. Only studies reporting actual teaching or learning activities with GenAI tools (e.g., LLMs, GPT-3.5, GPT-4, DALL·E) were included; opinion pieces, purely technical papers, and attitude-only surveys without an instructional intervention were excluded. To ensure conceptual rigor, we integrated the CIMO-Logic framework [58] with Biggs’s 3P model of learning as an interactive system (Presage–Process–Product) [59]. Within CIMO, we specified: (a) Context—the learning objectives pursued by children and adolescents in classrooms and homes; (b) Intervention—GenAI-supported activities ranging from prompt-based writing support to project-based inquiry; (c) Mechanism—the core human–AI role patterns that structure interaction; and (d) Outcome—both positive and negative learning effects.

We then mapped these CIMO elements onto the 3P model: Presage refers to the factors that exist before learning occurs, such as learner characteristics and the instructional context. Process encompasses the interactions and activities that take place during learning, including the pedagogical designs and the roles played by students, AI, and educators. Product represents the learning outcomes, both intended and unintended. This integrated lens allowed us to trace how aims, activities, roles, and outcomes co-vary and to interpret GenAI’s effects systemically. Figure 2 illustrates the combined framework.

Guided by these frameworks, we address four research questions:

RQ1 

(Presage): What primary learning goals does GenAI support for children and adolescents, and which educational needs do these goals address?

RQ2 

(Process): What learning activities are implemented with GenAI?

RQ3 

(Process): How do AI, students, teachers, and parents collaborate, and what role patterns emerge?

RQ4 

(Product): What opportunities and risks in learning outcomes are reported when GenAI is introduced?

3.1. Search Strategy

Between May and July 2025, we searched for empirical studies involving 3–18-year-old learners who used GenAI tools (e.g., ChatGPT, DALL·E) in formal or informal learning settings. All references were managed in EndNote and coded in Excel. We designed targeted search strings to capture the educational impacts of GenAI on children and adolescents and queried Scopus and Web of Science.

Table 1. Systematic search query for GenAI in k–12/children education. Note: The asterisk (*) denotes truncation in database search syntax (e.g., “LLM*” retrieves “LLM” and “LLMs”).

Search Dimension	Search Query
Generative AI	“generative ai” OR “generative artificial intelligence” OR “GenAI” OR “large language model *” OR “LLM *” OR “ChatGPT” OR “chat generative pre-trained transformer” OR “GPT-4o” OR “GPT-4” OR “GPT-3.5” OR “AI-generated content” OR “AIGC” OR “AI-generated” OR “generative model *”
k–12/Children	“k–12” OR “children” OR “students”
Education Context	“education” OR “learning”
Application Scene	“school” OR “home” OR “family”
Exclusion	NOT (“higher education” OR “university” OR “college student *” OR “undergraduate *” OR “postsecondary”)

lists the full search strategy.

3.2. Selection Method

We restricted the corpus to English-language, peer-reviewed journal or conference publications from 2020 onward to capture the most recent developments in GenAI in children’s education. We excluded reviews, posters, abstracts, extended abstracts, and forward-looking essays. After automatic de-duplication, screening proceeded on titles, abstracts, and keywords. Inclusion criteria required: (a) empirical studies (quantitative, qualitative, or mixed methods); (b) participants aged 3–18; (c) authentic teaching/learning activities in formal or informal contexts that actually used GenAI (e.g., LLMs, ChatGPT, AIGC); and (d) peer-reviewed publication. Exclusion criteria removed: higher education/adult training; purely technical implementations or attitude-only surveys without real learning interventions; non-GenAI (e.g., traditional ITS, discriminative models); and non-empirical or non-peer-reviewed work. Cross-database de-duplication relied on DOI and metadata matching. Two reviewers independently screened all studies to ensure reliability. Detailed criteria appear in

Table 2. Inclusion and exclusion criteria.

Category	Specific Criteria
Inclusion Criteria	1. Explicitly focus on the application of generative artificial intelligence (e.g., Generative AI, LLM, ChatGPT, AIGC, etc.) in educational settings for children and adolescents aged 3–18 (including K–12, children, primary and secondary school students, etc.).
	2. Study participants are children/adolescents aged 3 to 18 (may include teachers and parents as collaborative roles, but must involve learning or teaching activities with young learners).
	3. The study is empirical and involves concrete teaching interventions, classroom/home/project-based practices, or learning activities.
	4. Published as a peer-reviewed English journal article, conference paper, or other formal scholarly publication.
Exclusion Criteria	1. Unrelated to generative AI in child/adolescent education, or focuses on higher/adult education, teacher training, or non-generative AI (e.g., traditional adaptive systems, scoring systems, discriminative models, knowledge graphs, etc.).
	2. Participants are not aged 3–18 (e.g., only university students, adults, teachers), or the study setting is not relevant to K–12.
	3. Only reports technical implementation or surveys/interviews, without actual teaching/learning activities (e.g., only scoring historical data, teachers using AI for lesson preparation).
	4. Non-empirical research: reviews, opinion pieces, short abstracts, conference posters, non-peer-reviewed literature, or book chapters.
	5. Duplicate publications.

.

3.3. Selection Results

Our search in May 2025 identified 817 records (Scopus: 528; Web of Science: 289). After de-duplication and removal of items published before 2020, 605 remained. Title/abstract screening excluded 470, leaving 119 for full-text review. Of these, we removed 7 without K–12 participants, 16 that were not full empirical studies, 6 that were off-topic, 3 duplicates or substantially overlapping papers, and 1 non-English paper. Ultimately, 84 studies with MMAT (Hong et al., 2018) [60] scores ≥ 3 were retained for analysis (Figure 3).

3.4. Coding Scheme

To systematically analyze GenAI’s role in K–12 education, our coding scheme was aligned with the research questions and deepened theoretically in stages. As illustrated in Figure 4, our coding scheme builds on the 3P framework to map GenAI applications in K–12 education across presage, process, and product dimensions. Specifically: For learning objectives (RQ1), we first drew on Bloom’s taxonomy [61] and twenty-first-century skills frameworks [62], and—at the synthesis stage—incorporated generative learning theory (emphasizing deep learning through generating, organizing, and integrating rather than passive reception) and practice/situated orientations (learning as situated participation and meaning making) [25,63]. Accordingly, we treat higher-level goals as naturally subsuming lower-level performances so as to foreground learners’ active construction and transfer in complex contexts. For activity patterns (RQ2), we initially referenced UNICEF’s instructional activity framework [64] and then shifted to an activity-theoretic lens [65] to capture dynamic interactions among learners, tools, and communities. For human/AI roles (RQ3), we began with the three role paradigms—AI-directed, AI-supported, and AI-empowered [50]—and further drew on distributed cognition (cognition distributed across minds, bodies, and artifacts) [52], highlighting human–AI symbiosis and co-regulation and moving the unit of analysis from the individual learner to the human–AI ensemble. For learning outcomes (RQ4), we started from the OECD structure of knowledge, skills, and attitudes/values [55] and, informed by constructivist [66], situated [24], and sociocultural perspectives [67], reframed them as Epistemic, Practice, and Affective/Identity outcomes.

4. Results

Figure 5 illustrates the layered correspondence between learning objectives, learning activities, AI roles, and learning outcomes in GenAI-supported education. Studies focusing on language and literacy enhancement or STEM inquiry and practice most often employ AI-facilitated dialogic tutoring, generative iterative co-creation, or project-based problem-solving as learning activities. In these contexts, GenAI typically serves an AI-supported or AI-empowered role, with outcomes concentrated in Practice Outcomes and Epistemic Outcomes. In contrast, objectives related to Feedback Literacy and Self-Regulated Learning are frequently linked to formative feedback and adaptive scaffolding, where AI typically assumes an AI-directed role, primarily promoting outcomes related to affective and identity. Goals involving motivation and affect regulation, social–emotional skills and collaboration, or creativity and artistic expression are usually addressed through simulation and game-based learning, which feature diverse AI roles and lead to outcomes across Epistemic, Practice, and Affective/Identity outcomes. Overall, the type of learning goal largely determines the associated learning activities and the nature of AI participation, together shaping the final learning outcome. This reflects the dynamic and multifaceted pathways connecting goals, processes, and outcomes.

4.1. Learning Objectives

4.1.1. Language and Literacy Enhancement

We conceptualize language as a situated, dialogic, and generative practice [63,68,69]. As shown in Figure 6, 26 studies targeting language and literacy enhancement leverage GenAI to create immersive, interactive environments in which learners actively construct meaning through dialogue and feedback, which can be grouped into four categories: (1) Fostering deep, co-constructed understanding of texts. These studies addressed persistent challenges in traditional curricula, such as difficulties in understanding poetry, classical literature, or cultural symbolism. AI was used to help students engage with texts more deeply through visual and multimodal support [70,71,72,73,74]. (2) Providing scaffolded, interactive support. GenAI was employed to deliver personalized and leveled reading materials, helping to reduce obstacles such as comprehension difficulties, anxiety, or lack of confidence resulting from cognitive differences among students [75,76,77,78,79,80]. (3) Enabling a cycle of generation and revision. These studies used AI interventions to address common barriers such as limited teacher feedback, low student motivation for revision, or insufficient opportunities for creative expression [81,82,83,84,85,86,87]. (4) Creating more equitable opportunities for participation GenAI demonstrated distinct value by providing sustained, customized writing and reading support for students with dysgraphia, gender-related disadvantages, or rural and minority language backgrounds, thereby fostering confidence and promoting equitable development [88,89,90].

4.1.2. STEM Inquiry and Practice

In the context of STEM education, the primary learning objectives identified in recent research focus on engaging students in understanding abstract concepts, situating individualized learning, and deepening inquiry-based problem-solving practices. A central aim is to support students in actively constructing a more intuitive and in-depth understanding of complex scientific ideas. GenAI is employed to address these goals by supporting student engagement with challenging topics such as quantum theory, interdisciplinary knowledge transfer, and experimental design and hypothesis formulation—areas often limited by insufficient teacher resources and the lack of personalized instruction [91,92,93,94].

A second major objective is to situate the development of computational thinking and programming competence within authentic practices. Here, the target is to support students in refining their algorithmic reasoning, problem decomposition, and coding practices. The integration of AI-powered programming environments and intelligent feedback mechanisms serves to meet these goals, particularly by lowering entry barriers, providing personalized guidance, and ensuring timely feedback for novice learners—challenges frequently encountered in traditional computer science education [95,96,97,98].

Additionally, enhancing students’ critical thinking and real-world problem-solving practices remains a key objective. GenAI-supported interactive learning and analytical tasks are designed to promote critical analysis, logical reasoning, and the generative use of knowledge to authentic challenges in science and mathematics. These approaches also aim to address barriers such as the uncritical acceptance of AI-generated content, difficulties with complex texts, and a lack of self-efficacy, thereby supporting the development of autonomy and independent inquiry [99,100,101,102,103,104,105].

Overall, these studies address core challenges in STEM education, including resource constraints, lack of personalized support, and limited opportunities for complex, inquiry-based practices. The findings demonstrate that GenAI holds strong potential to deepen students’ engagement in practices of scientific literacy, critical thinking, and problem-solving.

4.1.3. Motivation and Affect Regulation

Studies in this strand use GenAI to stimulate students’ interest and intrinsic motivation, strengthen their sense of classroom belonging, and support emotional well-being and persistence [106,107,108]. They respond to well-documented problems in traditional instruction—low motivation, limited participation, and scarce affective support—by combining individualized AI interactions, timely feedback, and teacher facilitation to raise engagement, satisfaction, and confidence [109,110].

4.1.4. Creativity and Artistic Expression

This objective centers on inspiring ideas, enabling multimodal creation, and deepening reflection. Text-to-image systems are employed to enhance visual originality, learning motivation, and awareness of copyright and ethics [111,112,113,114]. Script and story generation translates children’s imagination into performable texts, improving reading fluency, narrative detail, and creative self-efficacy [115,116]. In design and problem-solving tasks, conversational scaffolds or CPS tutor agents are used to cultivate creative thinking and social responsibility, addressing the shortage of individualized guidance [117,118].

4.1.5. Social–Emotional Skills and Collaboration

Research in this category targets teamwork and social responsibility by building empathy and collective efficacy during negotiation, joint decision making, and supportive communication, thereby encouraging active engagement with social issues [119,120]. A second focus is emotion recognition and regulation in virtual environments, including the development of prosocial behavior and empathy [121,122]. A third line of research leverages AI to strengthen parent–child interaction and family collaboration, reinforcing emotional bonds, empathy, and parental involvement [121,123].

4.1.6. Feedback Literacy and Self-Regulated Learning

This objective aims to (1) strengthen students’ capacity to understand, act on, and self-regulate in response to AI-generated feedback [124,125,126,127]; (2) develop their ability to interpret, discriminate, and critically select among feedback from AI, teachers, and peers [128,129,130]; and (3) leverage AI to help teachers deliver just in time guidance for research and interdisciplinary projects, enabling students to clarify problems, structure ideas, and integrate knowledge [131]. Timely, personalized feedback offsets limited teacher resources, delayed responses, and insufficient individualization, promoting feedback uptake, autonomous revision, and sustained improvement. Collectively, these aims address long-standing deficits in personalized feedback, students’ motivation for self-regulation, and the scarcity of teacher guidance.

4.1.7. AI Literacy and Ethical Reasoning

This objective develops students’ understanding of how AI works, their ability to assess truthfulness, detect bias, and recognize the limits of creativity. It addresses deficits in technical understanding and resistance to deception [132,133,134]. Another thread examines ownership, data security, and academic integrity in collaborative knowledge-building contexts, responding to challenges in handling multiple viewpoints and academic responsibility [135,136]. Finally, through algorithmic auditing and reflective practice, students learn to critically evaluate AI decisions and societal impacts, reinforcing transparency, fairness, and accountability [137,138,139].

4.2. PROCESS: Learning Activities and AI Roles

4.2.1. RQ2 Learning Activities

(1) AI-facilitated Dialogic Tutoring: This activity type focuses on multi-turn question and answer exchanges with instant AI feedback, rather than on completing an entire task or project. In school settings, students repeatedly interact with AI assistants (such as ChatGPT, custom bots, and social robots) across English writing [85], reading comprehension [76,79,80,107,109], mathematics problem-solving [99,105], physics, quantum, and cognitive science [78,91,104,127], and computer programming [97,98,140,141]. These systems deliver personalized explanations, real-time examples, step-by-step reasoning, code demonstrations, assignment feedback, leveled explanations, and error diagnosis. Teachers employ tiered questioning, prompt calibration, and “virtual teacher” designs [92,142] to sustain an iterative cycle of adaptive probing, immediate feedback, and practice. In the home, parents coordinate with educational robots to tell stories and alternate questioning, thereby supporting shared reading and interactive learning [123]. Beyond answering, AI often acts as a dialogic coach or teachable agent that organizes a sequence of demonstration, explanation, questioning, and feedback, as illustrated by the APLUS mathematics environment and curiosity question training agents [6,102,128]. Several studies also require students to identify incorrect or fabricated AI outputs, which cultivates critical thinking and information discrimination skills [100,101,133].

(2) Generative Iterative Co-creation: Students act as creators who co-produce text, images, or multimedia artifacts with GenAI. Through continuous feedback and iteration, they deepen disciplinary understanding and enhance creative expression. In visual arts, learners employ Midjourney, DALL·E, and related text-to-image tools, repeatedly adjusting prompts to refine visual detail and aesthetics [111,112,113,114,115,116,132,143]; for example, teachers scaffold progressive prompting to render imagery from classical Chinese poetry [70]. Other studies use advanced models such as GANs and VAEs, allowing students to generate hybrid-animal images, match hand-drawn sketches, or explore latent-space interpolations via interactive play and creative exploration [134,136]. Large language models support literary writing, story crafting, and creative text generation; students iteratively revise content, adjust style, and sharpen narrative logic—often under teacher guidance to rework initial plots toward intended effects [84,86,139,144,145]. Mobile platforms further enable repeated sentence-level revisions to master grammar and expression [81]. In cross-disciplinary projects, students and teachers iteratively refine AI-generated scientific analogies to render abstract concepts more comprehensible (e.g., “the cell membrane is like a castle wall”) [94]. In mathematics and history, students use GenAI to iteratively generate and polish math stories, historical narratives, and concept maps [146], strengthening the understanding and application of abstract knowledge [72,88].

(3) Project-based Problem-Solving: Across studies, students used sustained, multi-turn interactions with GenAI to plan and execute inquiry- and project-based work. Typical activity sequences included: (a) co-defining authentic problems and drafting action plans (e.g., clean-water provision) [119]; (b) moving through the full research cycle in science courses—formulating questions, designing methods, and writing reports [131]; (c) posing and analysing real-world problems, constructing mathematical models, conducting scientific reasoning, and iteratively refining knowledge models [135,147]; (d) designing programming projects in graphical, conversational environments, with iterative AI-supported debugging, code structuring, and prototype refinement, including requirements gathering, function planning, and interface design for digital applications [95,96,117,148]; (e) undertaking scientific argumentation and critical inquiry—generating hypotheses, evaluating evidence, and revising claims and expositions—to strengthen the rigor of arguments [118]; and (f) coordinating collaboration in jigsaw-style tasks, where students aggregate individual inputs, generate prompting questions, and converge on shared experimental designs by specifying variables, materials, and procedures [93,108,120]. Some projects further embedded socio-technical critique, asking students to analyze recommender systems (e.g., TikTok) or prototype AI tools, thereby coupling creation, experimentation, and reflective ethics [137,138].

(4) Simulation and Game-based Learning: Learning is deliberately situated within immersive, narrative-rich virtual worlds, where students enact roles (players, designers, explorers) and engage in generative problem-solving around science, creative construction, and environmental themes. In Moon Story, an augmented reality system embedded in mythic narratives, learners communicate with fictional characters and complete astronomy and environmental missions [74]. In virtual construction spaces, AI acts as a design partner that dynamically joins discussions and optimizes ideas [71]. In science and outreach games, AI operates as a non-player character or assistant, providing stepwise hints, rule explanations, and plot advancement. Students progress through role exploration, collaborative reasoning, and strategic decision making [73,122,149].

(5) Formative Feedback and Adaptive Scaffolding: Immediate and personalized formative feedback, combined with adaptive support, promotes students’ self-revision, strategy adjustment, and continuous improvement. Across writing, reading, and mathematics tasks, students repeatedly submit texts, assignments, or answers to AI systems, receive concrete suggestions on error correction, structure, and language use, and then resubmit revised work, creating a cycle of practice, feedback, and renewed practice [75,77,82,83,87,90,110]. AI adjusts the specificity, depth, and difficulty of feedback based on students’ abilities and prior performance, enabling tailored prompts and tiered guidance. Examples include multi-dimensional advice for writing, stepwise guidance for reading or modeling, and targeted error localization [89,103,124,125,130,150], and LLM-based automatic question generation that dynamically constructs items, delivers immediate correctness judgments and targeted hints, and drives answer/feedback/revision loops; a light gamified interface (e.g., points, lives) is sometimes layered on only to sustain engagement [106]. Some studies also integrate social–emotional and interdisciplinary tasks, where AI provides focused evaluations and follow-up recommendations that help students refine learning strategies and strengthen self-regulation [121,129,151].

4.2.2. RQ3 AI Role

Building on the distribution in Figure 7, we detail the functional boundaries and evolutionary trends of the three AI role paradigms across activities.

(1) AI-directed, learner-as-recipient

This paradigm positions AI as the primary orchestrator of instruction: it sequences content and issues immediate feedback in a cue–response–feedback chain, while learners mainly comply with prompts and revise accordingly. Teachers act as quality and ethics gatekeepers, and—in younger cohorts—parents translate and soften AI feedback for children. The learner’s central task is thus to react to system cues in the expected manner.

A large subset of studies casts GenAI as a grader. During writing and reading assessments, AI delivers analytic diagnoses and scores across dimensions such as grammar, structure, and argumentation; teachers then prune or augment these prompts before returning them to students, who revise line-by-line, after which AI re-checks the latest draft [77,83,110,124,130]. In online courses, once teachers delimit the item bank, AI auto-generates questions and grades responses; students rewrite only the segments flagged as erroneous [77,106]. For younger learners, parents receive AI’s scoring notes, rephrase them in age-appropriate language, and supervise revisions [89]. Although teachers and parents retain oversight and translation roles, AI continues to set the pacing and structure, leaving students to respond passively to its prompts.

A second stream configures AI as a just-in-time tutor. When students encounter obstacles in programming or laboratory work, they query the model and receive stepwise explanations or sample code; teachers monitor frequency and content via dashboards and intervene only when AI misinterprets a problem or students stall for too long [101,148,150]. In more complex creative tasks, AI pushes stage-based strategies and affective support, while expert teachers mostly review interaction traces and fine-tune prompt templates [98,118,140,142]. Parents document difficulties that arise during home writing practice and upload them so teachers and AI can calibrate the granularity and timing of subsequent support [90]. This interactional design preserves the behaviorist chain: AI decides when to cue; students execute; and teachers supervise.

In a third configuration, AI functions as a materials planner. Systems mass-produce concept maps, leveled texts, or gamified problem sets and predefine reading order and pacing; after teachers screen and correct these resources, classroom control is largely ceded to AI. Students primarily browse, rate, or “unlock” these materials [76,133,146], or judge the correctness of AI-generated solution explanations post hoc, with teachers supplying clarifications as needed [105]. In studies on question-asking and analogy making, research teams pre-vetted GPT-3-generated prompts and analogies before importing them wholesale into instruction, after which students worked through them with scaffolded discussion [6,94].

(2) AI-supported, learner-as-collaborator

In this paradigm, under a distributed-cognition lens, learning activities are accomplished through the coordination of humans, AI, and artifacts; AI acts as an external cognitive partner providing dynamic scaffolding—continuously gathering learner input through dialogue, adapting examples and prompts in real time, and delivering personalized feedback. As competence grows, support fades, with control shifting from AI-led delivery to human–AI co-construction and learner self-regulation [52,152].

First, GenAI serves as a cognitive scaffold. By leveraging conversational interfaces, the system collects student input in real time, adaptively modifies examples and prompts, and integrates support functions such as simplified language, summarization, and length control within the same workspace. This enables interactive, immediate feedback, after which the AI gradually recedes as the task is completed [97,102,127].

Second, GenAI takes on the role of co-creator. For instance, ChatGPT can automatically generate and update concept maps that students then edit, enabling iterative refinement of external representations within a shared semantic space [146]. When students describe scenes in text, the system instantly generates corresponding images and allows further modification, supporting multimodal co-creation of learning content [72]. In writing support scenarios, AI provides immediate, personalized feedback after each draft and dynamically tailors subsequent suggestions based on students’ responsiveness, fostering cycles of collaborative authorship [84,87].

Third, GenAI functions as a reflective evaluator. By continually tracking students’ help-seeking behaviors and programming dialogues, the system automatically calibrates the granularity of its prompts and generates visual or textual feedback to help students monitor their progress and adjust learning strategies, thus enhancing metacognitive self-regulation [103,141,143,149]. In this process, teachers pre-author scaffolds for potential prompts and consciously relinquish some classroom control to uphold student agency [109]. From a distributed cognition perspective, GenAI is further positioned as an “external cognitive partner”, providing just-in-time support in inquiry-based tasks while ensuring that learners retain primary decision-making authority [93].

In sum, under this collaborative mode, GenAI moves beyond being a mere conduit for information. It serves as a real-time scaffold, co-creator, and reflective guide, collectively advancing a genuinely learner-centered and personalized educational process. If prompts and sequencing are preset, interaction regresses to AI-directed; when learners co-set evaluation criteria, delegate subtasks to AI, and coordinate with peers/teachers, the work approaches AI-empowered.

(3) AI-empowered, Learner-as-Leader

In the AI-empowered, learner-as-leader paradigm, learner agency is central; GenAI is configured as an augment to human intelligence within a negotiated human–AI partnership that enables co-regulated, distributed cognition [52]. From this perspective, education is a complex adaptive system in which effective learning outcomes emerge from the dynamic collaboration of multiple agents—learners, teachers, information sources, and technologies [153].

Within this framework, students actively set prompts and control content generation, while AI, functioning as an autonomous collaborator, continuously analyzes learner preferences, behaviors, and project progress to adaptively optimize its models for individual needs [70,96,108,132,144]. Teachers play a critical role as facilitators and supporters, working alongside AI to analyze learning data, refine creative or experimental processes, and provide timely guidance for topic selection, content iteration, and project reflection [112,113,116,136]. In some cases, educators employ GenAI to generate transparent and interpretable learning analytics reports, allowing for real-time adjustments to instructional strategies and more precise personalized interventions [122,151].

It is noteworthy that only a few studies have begun to explicitly position GenAI as an augmentation of human intelligence and to treat human–AI relations as a negotiated partnership; for example, by integrating ChatGPT into a Knowledge Building pedagogy to foreground a “human–AI partnership” and, under teacher shaping and guidance, encouraging students to mindfully configure that partnership and reflect on how AI is incorporated into learning [135].

Overall, these studies collectively show that AI, together with learners and educators, facilitates information exchange, iterative feedback, and collaborative knowledge building. Although most cases have not yet reached full-fledged hybrid intelligence, they are advancing along a path of co-regulation. This work lays the empirical groundwork for future learner-centered, adaptive, and human–AI symbiotic models of intelligent education.

4.2.3. PRODUCT: RQ4 Learning Outcomes

As shown in Figure 8, studies most frequently report affective and identity outcomes, followed by practice outcomes and epistemic outcomes. Figure 9 further summarizes the key opportunities and risks identified across these three theoretical domains.

4.2.4. Epistemic Outcomes

Grounded in constructivist learning theory, we define epistemic outcomes as the learner’s active reconstruction of beliefs about knowledge, evidence, and argumentation when encountering new information and experiences [66,154]. The core here is not the linear accumulation of facts, but a transformation in modes of reasoning, inferential practices, and critical judgment.

On the opportunity side, empirical studies consistently show that GenAI can promote cross-disciplinary knowledge building and conceptual deepening. Evidence spans STEM and the humanities, including gains in understanding physics [91,101,104,147] and programming [95,148,150], as well as improved reading and poetry interpretation [72,75,81,116]. More importantly, many studies report enhanced critical thinking [70,77,99,100,103,111,114,119,132,142,145,150], reflection, and problem-solving [70,128,141]. As a learning partner, AI provides high-quality concept maps [146], instructional analogies [94], and personalized feedback [128,130], creating deep learning experiences [95] and prompting reflection on socio-ethical implications of technology—thus seeding epistemic insight [117,132,135,138,151]. These gains are most often observed under teacher oversight, with adequate scaffolding, authentic tasks, and strong social negotiation.

Risks also surface when pedagogy is weak. If students engage only superficially with AI and fail to develop a critical understanding of how it works, deeper learning is unlikely. Studies document limited understanding of model training and datasets [112,122,135,136], which impairs students’ ability to detect errors or deceptive content [105,133] and leads to weak knowledge/analysis [109] or no significant improvement [98]. Readability issues [76], insufficient explanatory depth [91,92], and cognitive overload [73] can further impede knowledge construction and amplify over-reliance risks [94,119]. Put differently, “hallucinations” and inaccuracies are both hazards and—under guidance—teachable moments to cultivate critical evaluation. Accordingly, design should shift from answer-getting to treating AI output as a testable hypothesis within inquiry and co-construction.

4.2.5. Practice Outcomes

Grounded in situated learning and communities-of-practice perspectives, we define Practice Outcomes as the learner’s trajectory toward becoming a more competent participant in goal-directed community activities and dialogues [24]. This view treats learning not only as a cognitive act but as a social process of participation and interaction in authentic contexts, through which learners appropriate the discourse and repertoires of action of a given community (e.g., scientific inquiry).

On the opportunity side, studies show that GenAI can serve as a powerful medium that accelerates learners’ participation across diverse practice communities. First, AI provides strong scaffolds for core practices, markedly improving the quality and conventionality of learner outputs—for example, enhancing grammatical accuracy and argumentative coherence in writing [86,87,145], and supporting problem-solving and experimental design in science and mathematics [77,88,93,103,147]—thereby helping learners meet baseline practice demands more efficiently. Second, AI fosters generative practices: it supports creative writing [135], guides user-centered design thinking [117], enables more autonomous programming [96], and helps generate arts and history content [72,113], encouraging learners to move beyond “the right answer” toward more original exploration and construction. Finally, AI strengthens the social dimension of practice by facilitating peer collaboration [103,112] and parent–child dialogue [71,89]. Notably, these gains tend to depend on structured pedagogical design such as adaptive questioning [127,128], error analysis [77,103], or project-based learning [108].

However, bringing AI tools into practice communities also carries risks—especially when use is decoupled from the community’s core practices. A central risk is an illusion of participation: learners may appear to complete tasks while treating AI as an answer generator, failing to internalize the practice skills of the community. This shallow interaction shows up when some students make no substantive revisions after receiving AI feedback [115], or when teacher feedback is taken up more than AI feedback [130]. The deeper consequence is a drag on core capability development: over-reliance on AI for brainstorming or direct reuse can dampen originality and depth of creation [117,135,150]; in some contexts, traditional instruction even outperforms AI-integrated courses on writing and vocabulary [85]—underscoring that integration is not automatically beneficial. When AI use lacks clear scaffolding, when students’ tool-use competence is overestimated [127], or when tasks drift away from the complex, dynamic demands of authentic communities, GenAI can become an obstacle rather than an aid to practice.

4.2.6. Affective and Identity Outcomes

Grounded in sociocultural and identity perspectives, we define Affective and Identity Outcomes as the evolution of learners’ self-understanding, motivation, confidence, and capacity for purposeful action within a learning environment [25,67]. This lens attends not only to affective experience per se, but to the construction of a learner identity and the enactment of agency within a community of learning.

Empirically, GenAI is frequently associated with marked gains in affective engagement and motivation. Numerous studies report higher enthusiasm, participation, curiosity, and satisfaction after AI use [6,70,74,81,87,91,92,101,104,106,113,114,116,119,131,141,142], alongside a general preference for AI tools as efficient and convenient [99,111,124,127,135,140]. These affective gains often translate into identity-level shifts: students’ self-efficacy and domain confidence increase—particularly in programming [95], mathematics [77,103], and writing [86]—with corresponding reductions in learning anxiety [80,141]. Roles are also reshaped: learners move from passive consumers of knowledge to active creators of learning experiences [116], strengthening autonomy and agency [85,147]. At the relational level, AI can foster more positive peer affect and teamwork [107,120] and is positively appraised by parents and teachers for enabling parent–child dialogue and instructional support [71,121,123]. Teacher presence and guidance [82,107], together with personalization in AI tool design [97], are key enabling conditions.

That said, GenAI’s influence on affect and identity also carries risks. Some studies document confusion, frustration, and other negative emotions driven by mistrust, technical limits, or unmet expectations [78,87,93,98], and even declines in self-efficacy [140]. Motivation gains are not guaranteed: several studies find no significant effects [109] or only short-lived novelty bumps that taper over time [82,122]. These risks have multiple sources: uneven feedback quality [124], interaction designs that induce excessive cognitive load [84], and technical constraints (e.g., imperfect speech recognition, misleading “hallucinations”) [78,133]. Learner differences—cultural background, personal interests [129,132]—and concerns about academic integrity [85] further shape uptake. Parents and teachers likewise raise questions about age appropriateness, ethical risk, and the costs of integration [103,116,121,123,134]. Without deliberate attention to these factors, AI integration can heighten learning anxiety—or slide into a new form of technological dependence and passivity.

5. Conclusions

This systematic review synthesizes 84 empirical studies (2020–2025) on GenAI in K–12 and employs the 3P model (Presage–Process–Product) to examine learning objectives, activity patterns, AI role paradigms, and outcomes. We make three contributions: (i) an empirically grounded goal–activity–role mapping that traces how objectives shape activities and role allocation and, in turn, outcomes; (ii) a theory-forward reframing of outcomes—Epistemic, Practice, Affective/Identity—together with a role account that moves beyond “AI-as-tool” toward human–AI ensembles; and (iii) practice-ready guidance for responsible classroom adoption.

First, this review identified seven main categories of learning objectives: language and literacy enhancement, STEM inquiry and practice, motivation and emotional regulation, creativity and artistic expression, social–emotional skills, collaboration, feedback literacy and self-regulated learning, and AI literacy and ethics. Taken together, these objectives reflect a situated and generative view of learning: GenAI is used to create dialogic, multimodal environments that support meaning co-construction and cycles of generation–revision. Crucially, GenAI helps move beyond “basics-then-application” sequencing by enabling students to engage directly in complex, authentic problem-solving and creative production. These findings align with Marzano [30], who highlight GenAI’s promise for personalization and innovation in k–12 education. However, our review provides a more granular classification of objectives, multidisciplinary empirical evidence, and a clearer mapping from goal to activity, thus offering a more actionable reference for curriculum and instructional design.

Second, the role of AI in the classroom has shifted from that of information provider to collaborative partner and increasingly forming an integrated human–AI cognitive system. Based on the classic role framework proposed by Ouyang and Jiao [50], we find that while traditional AIEd research has centered on AI-directed scenarios, in which the system orchestrates the instructional process, positioning AI as a tool, our review documents a clear movement toward collaborative, empowerment configurations or even the emergence of augmentation of human intelligence, as seen in emerging roles such as co-author, reflective coach, and augmentative partner. This shift may be driven by the conversational and generative nature of GenAI systems, and allows control over learning tasks to be more equally shared.

This evolution is reflected in the alignment between activities and AI roles. AI-directed roles typically support structured activities with immediate feedback, whereas AI-supported roles are common in collaborative knowledge co-creation. AI-empowered roles, found in open-ended generative activities, emphasize learner agency and human–AI co-regulation. That said, we caution that simply adopting an “AI-empowered” approach does not automatically guarantee superior or more holistic learning outcomes. The benefits of granting learners greater leadership depend on activity design (for example, how much room is provided for student-driven decision-making and inquiry), as well as the scaffolding and oversight provided by teachers and parents. From a complex adaptive systems perspective, these shifts are enacted through feedback loops among students, teachers/caregivers, and GenAI: models provide context-aware prompts and suggestions; adult educators calibrate task scope, ethics, and related parameters; and learners accept, adapt, or challenge the guidance offered by AI, thereby shaping the system’s subsequent responses. As a result, agency is dynamically reallocated across parties rather than being vested in any single actor. Determining the specific conditions under which a relationship with AI as a teammate can reliably evolve into a co-regulated hybrid-intelligence model is an important avenue for future longitudinal research.

With regard to learning outcomes, our analysis moves beyond the traditional OECD framework of knowledge, skills, and attitudes/values, reframing them—in line with constructivist and sociocultural learning theories—as Epistemic, Practice, and Affective/Identity outcomes. Our findings confirm that GenAI enhances motivation and self-efficacy, consistent with Ali et al. [134] previous work. Notably, Affective and Identity outcomes were the most frequently reported, highlighting GenAI’s powerful role in strengthening learner agency and reshaping student identity from passive consumer to active creator. However, this reframing also clarifies GenAI-specific risks within a theoretical context: “hallucinations” pose a direct threat to Epistemic Outcomes by undermining knowledge construction, while the risk of an “illusion of participation” can hinder the development of genuine Practice Outcomes if students use AI as a mere answer generator. These findings suggest that GenAI is fundamentally reshaping the profile and risk structure of learning, moving beyond the challenges documented in prior AI literature.

Practically, these insights demand a shift in instructional scaffolding and safeguards. We outline three recommendations for responsible adoption: (1) Teacher professional development must focus not only on technical skills but on designing tasks that foster authentic disciplinary practice and critical epistemic inquiry. This includes orchestrating human–AI collaboration that avoids superficial engagement. (2) Feedback literacy for both students and teachers should be built around a calibrate–critique–act cycle, treating AI output as a testable hypothesis. This reframes AI feedback as an opportunity to deepen practice rather than as a final truth to be received. (3) Ethical guardrails remain essential, including age-appropriate privacy, routine audits for bias and hallucination, and clear integrity protocols, which collectively protect learners’ affective well-being and trust.

Limitations and Future Work

However, this review has several limitations. Most included studies focused on short-term interventions, making it difficult to assess the sustained impact of GenAI on long-term cognitive, skill, and attitudinal development. In addition, few studies systematically assessed algorithmic bias or privacy risks, which may lead to an underestimation of negative outcomes. Our review also did not distinguish between formal and informal learning environments or among different age groups, which may mask the moderating effects of context and developmental level on the effectiveness of AI applications.

These gaps motivate several research directions: (1) Longitudinal and cross-cultural studies are needed to trace the long-term evolution of epistemic, practice, and identity outcomes across different developmental stages. (2) Research should investigate the micro-processes of pedagogical design, using co-design approaches with all stakeholders. This includes identifying scaffolding techniques that foster a co-regulated human-AI partnership and understanding the critical dynamics of when AI should provide support versus handing off to peers and teachers. (3) Developing and validating new assessment paradigms capable of capturing the complex, process-oriented learning identified in this review, particularly the development of epistemic outcomes and shifts in learner identity.