Despite promising effectiveness in isolated applications, AI technologies in elementary STEM education remain fragmented across disconnected systems, with only 15% of studies addressing integrated STEM learning, revealing eight critical gaps—including developmental inappropriateness, infrastructure barriers, and equity disparities—that prevent AI from achieving its transformative potential for young learners.
Objective
This systematic review aimed to comprehensively examine AI applications in elementary STEM education (grades K-6, ages 5-12) by addressing four critical research questions: identifying what AI technologies are currently deployed and how they are distributed across grade levels and subjects; evaluating evidence for their effectiveness in supporting learning outcomes; determining key implementation challenges and barriers; and identifying gaps in current research and practice that must be addressed to realize AI's potential. The review sought to move beyond fragmented understandings of individual technologies toward a comprehensive view of AI's current and potential roles in supporting young learners' STEM education, with particular attention to how AI supports or hinders integrated learning across Science, Technology, Engineering, and Mathematics disciplines.
Methods
The researchers conducted a systematic review following PRISMA guidelines, searching six major academic databases (Google Scholar, PubMed, IEEE Xplore, ACM Digital Library, ERIC, and PsycINFO) between September and October 2025. The search targeted publications from 2020-2025 to capture post-pandemic developments, while including earlier seminal works when relevant. The comprehensive search strategy encompassed eight AI technology categories: intelligent tutoring systems and conversational AI, automated assessment and feedback systems, learning analytics and predictive modeling, computer vision for engagement monitoring, multimodal sensing and biometric analysis, adaptive content generation, educational robots and embodied AI, and extended reality with AI enhancement.
Studies were included if they addressed elementary-age populations (K-5/6, ages 5-12), investigated one or more AI technologies applied to educational contexts, examined STEM learning outcomes in individual subjects or integrated approaches, employed empirical methods including randomized controlled trials, quasi-experiments, design-based research, case studies, or systematic reviews, and were published in peer-reviewed venues. The initial search identified 3,847 records, which after removing duplicates and screening resulted in 513 full-text articles assessed for eligibility. The final synthesis included 258 studies meeting all inclusion criteria.
For each included study, researchers extracted population characteristics, AI technology type and specific features, STEM domains addressed, implementation context, outcome measures, effect sizes where reported, implementation challenges, and identified limitations. Quality was assessed using the Mixed Methods Appraisal Tool (MMAT) version 2018, with two reviewers independently appraising 20% of studies (achieving substantial inter-rater reliability with Cohen's κ = 0.78). Thematic synthesis identified patterns across studies, organized by AI technology category while attending to cross-cutting themes such as developmental appropriateness, equity considerations, teacher roles, and infrastructure requirements.
Key Findings
The systematic review revealed a rapidly evolving but fragmented landscape dominated by intelligent tutoring systems and conversational AI (45% of studies), followed by learning analytics (18%) and automated assessment (12%). Emerging technologies including computer vision (8%), educational robotics (7%), multimodal sensing (6%), and AI-enhanced extended reality (4%) showed growing but limited deployment. Geographic concentration was stark, with 90% of studies from North America (42%), East Asia (28%), and Europe (20%). Grade-level coverage showed bias toward upper elementary grades 3-5 (65%) versus K-2 (35%), while mathematics dominated subject focus (38%), followed by computational thinking/coding (26%), science (22%), and engineering (14%). Critically, only 15% of studies attempted integration across multiple STEM domains.
Effectiveness evidence varied substantially across technologies. Conversational AI showed moderate effectiveness with effect sizes of d=0.45-0.70 where reported, particularly for mathematics tutoring and inquiry-based science learning. However, only 34% of studies reported standardized effect sizes, limiting quantitative synthesis. Learning analytics demonstrated promise for early warning systems but exhibited bias in predictive models, performing poorly for students whose learning patterns differed from training data. Automated assessment tools showed strength in evaluating procedural knowledge but struggled with open-ended responses and conceptual understanding. Computer vision applications raised significant privacy concerns and reliability issues with young children, while cultural bias in training data presented critical limitations. Educational robots demonstrated potential for learning-by-teaching paradigms and integrated STEM projects, with notable success in robotics ecosystems combining biology, engineering, mathematics, and computational thinking. However, high costs ($3,000-3,500 per unit for advanced systems), maintenance burdens, and proprietary platforms limited widespread adoption. Extended reality applications showed promise for spatial and experiential STEM learning but faced substantial barriers including hardware costs, motion sickness concerns, age-appropriate content limitations, and technical expertise requirements.
The analysis identified eight critical systemic gaps preventing AI from achieving its potential: (1) fragmented ecosystem where AI technologies operate in isolation without interoperability standards, reinforcing rather than overcoming subject silos; (2) developmental inappropriateness, with one-size-fits-all approaches ignoring profound differences between K-2 and grades 3-5 learners; (3) infrastructure barriers requiring high-speed internet, modern devices, and technical support unavailable in many schools, creating technological redlining; (4) privacy and ethical void with extensive data collection from minors occurring without comprehensive governance frameworks; (5) limited STEM integration with mathematics, science, and coding tools operating separately without cross-disciplinary connections; (6) equity and access disparities concentrating advanced AI in well-resourced schools while under-resourced schools struggle; (7) teacher marginalization through black-box algorithms and overwhelming dashboards that disempower educators; and (8) narrow assessment focus excelling at procedural knowledge but failing to capture creativity, collaboration, and conceptual understanding.
Implementation challenges transcended individual technologies. Infrastructure requirements created compounding disadvantages for rural schools and those serving low-income communities. Elementary teachers, typically generalists rather than STEM specialists, reported feeling overwhelmed by technical complexity, with professional development focusing on tool operation rather than pedagogical integration. Privacy concerns regarding extensive data collection—from clickstream data to biometric monitoring—raised serious questions about student privacy, commercial use of information, and long-term implications of creating detailed digital profiles of young learners. The absence of interoperability meant each AI system operated in isolation with proprietary data formats, unique interfaces, and separate login credentials, creating fragmented experiences. Language barriers, cultural biases in AI training data, and assumptions about home technology access created additional obstacles for marginalized communities, threatening to widen rather than narrow achievement gaps.
Implications
This research contributes significantly to understanding AI's role in elementary STEM education by revealing that current implementations fail to address the integrated, developmentally appropriate, and equitable learning experiences that young students need. The concentration of AI deployment in isolated subject areas contradicts the fundamental premise of STEM education—that these disciplines are interconnected and mutually reinforcing. The finding that only 15% of studies address integrated STEM learning highlights a critical missed opportunity, as AI technologies could theoretically support the cross-disciplinary connections that human teachers struggle to facilitate.
The review's identification of eight systemic gaps provides a framework for understanding why AI's promise remains largely unrealized despite technological sophistication. These gaps are interconnected: infrastructure barriers prevent equitable access, developmental inappropriateness undermines effectiveness for younger children, privacy concerns create legitimate hesitation among parents and educators, fragmented ecosystems reinforce subject silos, teacher marginalization reduces implementation fidelity, and narrow assessments drive curriculum narrowing. Addressing these challenges requires coordinated solutions rather than piecemeal technological fixes.
The paradigm shift toward workforce preparation and AI-driven educational transformation represents both an opportunity and a concern. Models like Alpha School, Unbound Academy, and microschools demonstrate how AI can reshape entire educational experiences, potentially supporting interest-driven, career-connected learning from elementary grades. However, Alpha School's $40,000-$65,000 annual tuition exemplifies how this transformation risks exacerbating inequities rather than addressing them. The review suggests that realizing AI's democratizing potential requires explicit attention to equity, infrastructure development, and teacher empowerment.
For practitioners, the findings suggest that successful AI integration requires treating AI as one component of effective STEM instruction rather than a replacement for human expertise. The evidence supports teacher-centered implementations that enhance rather than replace educator agency, with professional development focusing on pedagogical integration rather than tool operation. Schools should prioritize pilot programs with careful evaluation to identify which technologies provide genuine value versus those that merely add complexity, while recognizing that infrastructure and teacher preparation must precede large-scale AI deployment.
Limitations
Several important limitations affect interpretation of findings. The restriction to English-language publications in academic databases potentially excludes relevant work from non-English-speaking regions where AI in education research advances rapidly. This language bias particularly underrepresents innovations from countries not captured in the predominantly North American, East Asian, and European corpus. The heterogeneity of study designs, outcome measures, and implementation contexts prevented formal meta-analysis, with only 34% of studies reporting standardized effect sizes limiting quantitative synthesis. Wide variation in intervention durations (from single sessions to full academic years), sample sizes (from n<20 pilots to district-wide implementations), and outcome assessments makes direct comparisons challenging.
The 2020-2025 time window, while capturing recent post-pandemic developments, may miss foundational work that continues to influence current implementations. Additionally, the rapid pace of AI advancement means some reviewed technologies may already be superseded by newer approaches not yet documented in peer-reviewed literature. Publication bias likely inflates apparent effectiveness of AI interventions, as studies reporting null or negative results may be underrepresented. The concentration of research in well-resourced contexts limits generalizability to diverse global settings, while the predominance of studies in upper elementary grades (65%) and mathematics (38%) may not reflect AI's full potential or challenges across all elementary STEM contexts.
Quality appraisal revealed common concerns including limited sample sizes in pilot studies (38% with n<50), short intervention durations (45% under 4 weeks), lack of control groups in quasi-experimental designs (52%), and inconsistent effect size reporting. These quality considerations informed the narrative synthesis, with higher-quality studies given greater weight in drawing conclusions. Finally, the focus on empirical studies may undervalue important theoretical contributions and design frameworks lacking formal evaluation but offering valuable insights for the field.
Future Directions
The review identifies urgent research priorities organized into several interconnected areas. First, addressing the fragmented ecosystem requires systematic investigation of integration architectures enabling different AI technologies to work together coherently, with development of interoperability standards and shared pedagogical frameworks supporting cross-disciplinary STEM learning. Second, the developmental inappropriateness of current systems demands longitudinal studies tracking how children of different ages interact with and benefit from various AI technologies, with explicit attention to designing grade-appropriate interfaces, interaction patterns, and cognitive demands differentiated for K-2 versus grades 3-5 learners.
Third, the privacy and ethical void requires immediate development of comprehensive frameworks appropriate for educational contexts involving minors, including clear governance for data collection, storage, analysis, and protection, transparent communication with parents, and age-appropriate consent processes. Fourth, addressing equity disparities necessitates research examining how AI-driven educational transformation can be implemented equitably across diverse elementary settings, particularly in under-resourced rural and urban schools, with explicit investigation of language barriers, cultural biases, and assumptions about home technology access.
Fifth, evaluating emerging school-wide AI models (Alpha School, Unbound Academy, microschools) through rigorous longitudinal studies can establish their scalability, cost-effectiveness, and equity implications compared to traditional approaches. Sixth, investigating AI-powered career discovery systems for elementary students and their impact on STEM engagement and motivation could inform the paradigm shift toward workforce preparation, with careful attention to whether career-connected learning produces better outcomes than traditional academic preparation.
The authors propose a novel experimental framework examining AI-Human collaborative learning ecosystems through a four-condition randomized controlled trial across 40 schools and 1,200 students, comparing AI-Human collaborative, AI-only, human-only, and business-as-usual conditions. This research aims to establish evidence-based frameworks for optimal AI-Human task allocation, culturally-responsive AI development, and career-connected learning pathways. Expected contributions include theoretical development of AI-Human Collaboration Theory, Cultural AI Theory, and Career-Connected Learning Theory; methodological advances in mixed-methods AI evaluation and longitudinal career outcome assessment; and practical protocols for implementing AI-human collaboration, developing culturally-responsive AI systems, and creating career-connected curricula.
Perhaps most critically, future research must move beyond isolated technology studies to examine AI as part of complex educational ecosystems, investigating interactions between multiple AI systems, human teachers, diverse learners, and varied contexts. Research methodologies must evolve beyond traditional experimental designs isolating single variables toward mixed-methods approaches combining quantitative outcome measures with qualitative process data to understand not just whether AI systems work, but how and why they succeed or fail in real elementary classrooms.
Title and Authors
Title: "Artificial Intelligence in Elementary STEM Education: A Systematic Review of Current Applications and Future Challenges"
Authors: Majid Memari (Department of Computer Science, Utah Valley University) and Krista Ruggles (School of Education, Utah Valley University)
Published On: November 6, 2025 (arXiv preprint)
Published By: arXiv (preprint server) - Submitted to peer-reviewed journal