Decade-Long Review of AI in STEM Education Charts a Shift From Tutoring Bots to Inquiry-Driven Learning

A ten-year sweep of the academic literature on artificial intelligence in STEM education has surfaced a clear inflection point: the field is moving away from rigid intelligent tutoring systems toward tools that cultivate inquiry-based learning and computational thinking. The finding comes from a bibliometric analysis of 242 publications spanning 2015 to 2025, submitted to arXiv on June 10, 2026, and accepted for presentation at the ISLS26 conference.

Rather than proposing a new model or classroom pilot, the paper takes stock of how researchers have studied AI's role in science, technology, engineering, and math instruction over the past decade. By mapping citation patterns and thematic clusters across nearly a quarter-century's worth of publications compressed into a ten-year window, the authors trace a shift in the underlying philosophy of educational AI: early systems were built to deliver structured, rule-based tutoring, while more recent work leans on large language models to support open-ended exploration and problem-solving skills.

From Scripted Tutors to Scaffolded Inquiry

The paper's central argument is that the rise of LLMs has changed what AI is expected to do in a classroom setting. Traditional intelligent tutoring systems were designed to walk students through predefined problem paths, correcting errors against a fixed answer key. The literature reviewed shows that emphasis fading in favor of systems that support students as they formulate questions, test hypotheses, and reason through computational problems on their own terms.

According to the authors, the throughline connecting these approaches, old and new, is scaffolding. Their analysis concludes that AI's most durable contribution to STEM education is not automation of instruction itself, but the reduction of the cognitive threshold required to grasp complex material. In other words, well-designed AI support does not replace the hard work of learning; it removes unnecessary friction so students can engage with harder concepts sooner.

Why the Distinction Matters

This is a meaningful distinction for anyone building or funding educational AI tools. A system optimized purely for answer accuracy and step-by-step correction is a fundamentally different product than one optimized to prompt curiosity and structured reasoning. The bibliometric evidence suggests the research community's priorities have already tilted toward the latter, which has implications for how ed-tech companies pitch their LLM-based tutoring products and how institutions evaluate them.

The paper does not report experimental results of its own; its contribution is synthesis rather than new empirical testing. Its value lies in giving educators, developers, and policymakers a decade-spanning map of where the field has been and where its center of gravity is heading. With acceptance at ISLS26, a venue dedicated to the learning sciences, the findings are likely to feed directly into ongoing debates about how generative AI should be integrated into science and math curricula, particularly around the question of whether scaffolding by AI ultimately deepens understanding or simply defers it.