The historical arc of the field is longer than most people realize. Long before ChatGPT captured headlines, rule-based intelligent tutoring systems such as SCHOLAR and GUIDON were already supporting science and mathematics instruction in the 1970s and 1980s, grounded in cognitive science models of how students learn. The 1990s and 2000s saw cognitive tutors and early adaptive platforms migrate from laboratory experiments into real educational practice, while learning analytics slowly began informing instructional decisions. The 2010s brought data-driven adaptive systems, machine learning techniques, and the massive open online course boom, which collectively expanded AI-supported learning environments across the globe. Since roughly 2020, however, generative AI and large language models have marked a genuine turning point, and the publication record shows it dramatically: between the early 1990s and about 2018, research output in this area remained minimal and sporadic, but from 2020 onward the number of publications surged sharply, confirming that AI in science education has become one of the fastest-growing research domains in the learning sciences.

Beneath that headline growth, however, the review uncovers a strikingly lopsided geography of research effort. Of the 80 studies analyzed, 51 percent were conducted in higher education contexts, while only 24 percent took place in K-12 settings and a mere 8 percent in teacher education. This concentration makes structural sense—universities have the organizational flexibility and research infrastructure to pilot and evaluate AI-based approaches—but it means the foundational stage of education, where scientific literacy and digital competencies first take root, remains largely unstudied. Compulsory schooling, where millions of children first encounter physics, chemistry, and biology, is precisely where we know least about how AI tools behave, how teachers adapt them, and how students of different ages respond to algorithmically mediated learning.

The pattern of curriculum integration mirrors this imbalance. Institutional and program-level integration dominated the reviewed literature at 41 percent, with full integration across institutional, program, or policy levels accounting for 55 percent of studies when categories were aggregated. Course and module-level integration followed at 26 percent, while lesson and classroom-level integration trailed at just 11 percent. The review interprets this as a predominantly top-down pattern of implementation: AI is being embedded within broad curricular structures and administrative frameworks rather than emerging organically from individual teachers experimenting in their own classrooms. Notably, the dominance of institutional-level integration aligns with the concentration of studies in higher education, where program-level curricular decisions are more feasible. In K-12 settings, by contrast, integration appears localized and fragmented—isolated lessons or extracurricular use—suggesting structural barriers that limit systemic adoption in compulsory education. Policy and system-level integration remained rare at 14 percent, indicating that even as AI colonizes institutional curricula, it has barely penetrated the governance structures and assessment practices that shape educational systems as a whole.

What are schools and universities actually doing with AI? The answer, overwhelmingly, is personalization. Personalization and adaptive learning emerged as the dominant pedagogical application, appearing in 59 percent of the reviewed studies. Simulation and modeling came a distant second at 21 percent, reflecting AI’s capacity to visualize scientific processes and support laboratory-oriented learning—generative tools such as ChatGPT have even been examined as virtual laboratory teaching assistants that help students design experiments, interpret data, and strengthen scientific reasoning. Inquiry support, content generation, feedback, tutoring, and teacher planning each accounted for small fractions of the remaining applications. This concentration matters because adaptive learning systems are among the most technologically complex and opaque forms of educational AI, relying on continuous collection and processing of learner data to tune instruction in real time. The very features that make them pedagogically attractive also make them the most demanding from a transparency and accountability standpoint.

That tension comes into sharp focus in the review’s ethical analysis, which yields perhaps its most striking findings. Transparency and explainability topped the list of ethical concerns at 25 percent of studies, followed by student agency at 21 percent and the teacher’s role at 13 percent. Taken together, student agency and teacher role account for 34 percent of all ethical considerations, revealing a substantial human-centered strain in the literature that foregrounds autonomy, professional identity, and the relational dynamics between human and artificial agents. The review argues this is no accident: ethical awareness in the field appears functionally connected to the technologies under study, with the opacity of adaptive algorithms directly driving the elevated concern for explainability. Yet the flip side is sobering. Access and equity drew only 6 percent of ethical attention, bias and fairness just 5 percent, and governance and policy a mere 5 percent.

The single most alarming number in the entire review concerns privacy. Just one study—1 percent of the corpus—addressed privacy and data protection as its principal ethical concern, despite the fact that nearly six in ten studies examined data-intensive personalization systems. The review offers several explanations for this blind spot. Privacy is frequently framed as a technical or legal compliance matter rather than a pedagogical concern, reducing its visibility in educational research. Regional differences in data-governance frameworks, such as the GDPR in Europe and FERPA in the United States, create inconsistent treatment of learner data across contexts. Ethics review processes for studies using commercial AI tools or secondary data may simply overlook privacy implications for learners. And the breakneck pace of generative AI adoption has outrun the scholarship meant to evaluate it, creating a temporal gap between technological innovation and research on data protection. Whatever the cause, the review concludes that the field has not yet adequately engaged with the data governance implications of its single most popular AI application.

To knit these fragmented threads together, Molnár proposes a multidimensional framework that conceptualizes AI integration along three interdependent dimensions: the type of AI technology involved, the level of curriculum integration, and the ethical focus of that integration. The framework complements established models such as TPACK, which describes the knowledge teachers need for technology integration, and SAMR, which characterizes increasing levels of task transformation. Its distinctive contribution is the explicit incorporation of ethics into a single analytical structure at the research-synthesis level. An exploratory statistical test within the corpus found only weak, non-robust support for the proposed interdependencies—the association between integration level and ethical engagement did not survive permutation testing—but the patterns remain suggestive. Studies with limited curriculum integration tended to engage with ethical issues only superficially, while more comprehensive integration more frequently came with explicitly stated ethical concern. The framework’s relationships are therefore advanced as propositions for testing in larger samples rather than established regularities, but they offer a practical planning tool: introducing a generative AI writing tutor into a secondary chemistry module, for example, requires systematic evaluation of technology type, integration depth, and ethical dimensions including transparency, agency, and data governance.

The review is candid about its own limitations, which is itself refreshing in a field prone to hype. The single-author design precluded formal inter-rater reliability, though a blinded intra-rater protocol—full recoding of all 80 studies after a minimum two-week interval—achieved 83 percent agreement with Cohen’s kappa of 0.83, indicating almost perfect agreement by conventional benchmarks. The reliance on Web of Science and Scopus excludes educationally oriented studies indexed elsewhere, the English-language restriction may sideline research traditions in East Asia, Latin America, and Central Europe, and the post-2020 concentration of the corpus limits longitudinal conclusions. Publication bias is also a real risk, since successful AI implementations are more likely to be published than failures, potentially producing an overly optimistic picture of what AI actually achieves in science classrooms.

For educators, the practical message is that top-down mandates do not automatically translate into effective classroom practice. The review calls for bottom-up approaches centered on teacher-initiated classroom pilots, collaborative communities of practice, and co-design methods that adapt AI to diverse contexts, supported by sustained professional learning that positions teachers as informed pedagogical decision-makers rather than passive implementers of prescribed technology. With only 8 percent of studies addressing teacher education, the evidence base for preparing educators remains thin, even as TPACK-based studies of science teachers report significant gaps in design competencies and widespread dissatisfaction with existing professional development. For policymakers, the picture is starker still: institutional practice is outpacing regulation, and the review urges frameworks that balance innovation with the protection of learner rights. The deeper conclusion is that AI in science education cannot be understood as a purely technological question. It is simultaneously a pedagogical, ethical, and societal transformation—and right now, the research community is studying the technology while the ethics, the governance, and the classrooms of compulsory education lag dangerously behind.

Subject of Research: Artificial intelligence integration in science education, including publication trends, curriculum implementation, and ethical considerations

Article Title: A multidimensional review of artificial intelligence in science education examining trends, curriculum integration, and ethical implications

Article References: Molnár, Z. (2026). A multidimensional review of artificial intelligence in science education examining trends, curriculum integration, and ethical implications. Discover Education, 5(1), Article 988. https://doi.org/10.1007/s44217-026-02194-2

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02194-2

Keywords: artificial intelligence, science education, systematic review, curriculum integration, adaptive learning, generative AI, ethics, data privacy, K-12 education, teacher education, bibliometrics, education policy