Research Article
Artificial Intelligence in Elementary Science Education: Mapping Research Trends, Educational Applications, and Emerging Challenges
Synthesis: Barokah, Hermita, and Putra (2026) map 103 Scopus-indexed publications from 2023 to early 2026 to describe how research on artificial intelligence is being conducted in elementary science education. They report rapid growth alongside weak consolidation: much of the corpus is general AI-in-education reviews, bibliometric work, or higher education studies, and only a smaller cluster addresses primary science directly. They sort the literature into five application clusters, from generative lesson design to educational robotics, and find outcomes increasingly measured affectively and regulatorily while epistemic outcomes such as evaluating evidence remain thin. Their contribution is a three-dimensional framework (cognitive-epistemic, instructional-interactional, systemic-ethical) plus open challenges spanning longitudinal evidence, teacher readiness, assessment validity for young learners, equity, and ethical governance. The paper measures the shape of a literature, not the effects of any tool.
Key Findings
- The field is expanding faster than it is consolidating. The authors report a sharp rise in Scopus-indexed work linking AI, elementary education, and science learning across 2023 to 2026, yet much of the corpus is broad AI-in-education reviews, bibliometric studies, or higher education work, so elementary science remains "not yet a fully consolidated subfield."
- Five application clusters carry the field. Generative AI for lesson design, personalized and adaptive support, AI-based assessment, educational robotics, and AI literacy with teacher development each promise something different and fail differently.
- Evidence concentrates geographically. The Scopus summary names China, the United States, and Hong Kong as leading contributors, while rural, low-resource, Indigenous, and Global South contexts stay largely invisible, even though AI integration in primary science depends on infrastructure, language, and curriculum alignment.
- Outcomes broaden but stop short of epistemic goals. Cognitive, affective, regulatory, and participatory outcomes all appear, yet studies give less attention to how children evaluate evidence, separate observation from inference, construct models, or recognize the limits of AI-generated claims.
- Validity, not scoring accuracy, binds assessment. Upper elementary performance tasks yield partial explanations, drawings, gestures, and emerging language; scoring that marks a developmentally productive response as simply wrong strips assessment of its formative value.
- Reviews outnumber classroom studies. Bibliometric analyses, systematic reviews, and conceptual papers far outnumber classroom-based empirical studies, and longitudinal designs specific to elementary science are rare.
How the map was built
This is a systematic mapping study, not an effectiveness review, and the authors keep the distinction throughout: the goal is to describe a literature's characteristics, distribution, and thematic structure, not to estimate an effect. The search ran in Scopus between January and March 2026 and covered publications from January 2023 to February 2026, a window chosen for the arrival of generative AI. After screening, the corpus held 103 publications coded across eight analytical dimensions, from science topic and AI technology to learning outcomes, research method, and emerging challenges.
The analysis was qualitative content analysis supported by descriptive mapping, with iterative author discussion in place of formal quality appraisal. That design explains the paper's character: it can say which technologies, topics, and theories recur, but not whether any improved children's science learning, and the authors say so. The growth claim is likewise a described pattern, not a curve: the paper prints no year-by-year counts.
Five application clusters, five failure modes
The clusters are the most usable part of the mapping. Generative AI appears mainly in teacher-facing work: drafting lessons, producing multiple explanations of one concept, adapting tasks to reading levels, building local environmental examples. Its risk is scientific error at the moment a teacher is shortest on subject expertise, since elementary teachers often teach across subjects. Personalization and adaptive support promise differentiated pathways and immediate feedback, but the review warns against individualization detached from classroom discourse, since pupils need to compare evidence and negotiate explanations. AI-based assessment tackles the real problem of interpreting drawn, written, and modeled artifacts at scale. Robotics bridges computational ideas and physical phenomena, though engagement in a build challenge does not by itself produce understanding.
Across clusters the same condition recurs: value depends on how the teacher frames and mediates the task. The authors support that by citing other work, not by anything their coding produced.
Theories that survive the technology
The mapped studies lean on constructivist and sociocultural foundations, technology acceptance models, self-determination theory, and self-regulated learning frameworks, with human-centered AI and socio-technical governance entering recently. AI does not displace these theories but pressures them: constructivism now must account for children building knowledge alongside machine-generated explanations. The three-dimensional framework organizes that pressure; it is a proposal, not a tested model, and no study in the corpus evaluates it.
What this means for practice
- Instructors: treat generative AI as a planning partner and verify each explanation for scientific accuracy, developmental fit, and local relevance before it reaches children.
- Curriculum designers: specify the science concept and practice a task teaches, not just its STEM label; a robotics challenge may build collaboration while teaching nothing about force or ecosystems.
- Administrators: fund professional development that joins AI literacy to science pedagogy, including judging outputs and protecting student data, not generic prompt-writing training.
- Assessment designers: build scoring around partial, drawn, oral, and multimodal evidence, and keep teacher interpretation inside the loop rather than treating automated scores as the outcome.
- Policymakers: require transparent rules on student data, human oversight, and age-appropriate consent before elementary classrooms adopt these tools at scale.
Limitations
- No effectiveness evidence. The mapping codes publication characteristics and themes, not outcomes, so nothing here shows any application improving science learning.
- Single database, short window. Restricting to Scopus and to 2023 through early 2026 excludes regional journals and the longer history of AI in science education.
- Coding is qualitative and unregistered. No second-coder agreement statistic and no formal quality appraisal mean cluster boundaries are authorial judgments.
- Growth claims are unquantified. The sharp increase is asserted without year-by-year counts or trend statistics.
- Adjacent literature inflates the base. Many included studies concern STEM broadly, teacher education, or assessment in general, weakening inferences about elementary science.
Citation
Barokah, R. G. S., Hermita, N., & Putra, S. P. (2026). Artificial Intelligence in Elementary Science Education: Mapping Research Trends, Educational Applications, and Emerging Challenges. Journal of Studies in Elementary Education, 2(1), 31–46.