Artificial Intelligence, Sustainability, and the Development of Mathematical Thinking: A Theory-Grounded Scoping Review

Artificial intelligence (AI) tools are increasingly integrated into mathematics education, yet most reviews emphasize achievement rather than how AI shapes mathematical thinking. This scoping review mapped literature published between 2020 and 2026 on AI-supported mathematics learning through three cognition frameworks: APOS (Action–Process–Object–Schema), Sfard’s process–object duality and reification, and Conceptual Image theory. Searches were conducted in Scopus, Web of Science, ERIC, PsycINFO, Education Source, and IEEE Xplore, followed by duplicate removal and Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews (PRISMA-ScR)-aligned screening. Twenty-one peer-reviewed studies met inclusion criteria (18 empirical studies plus three theoretically oriented studies). Evidence growth accelerated after 2022, with most studies situated in secondary and higher education. Large language models (LLMs) and Intelligent Tutoring Systems (ITS) were the most frequently investigated modalities. Across studies, AI commonly supported theoretically inferred action-level execution and procedural management (APOS) via adaptive feedback, hinting, and stepwise scaffolding, and it often broadened learners’ conceptual images through multiple representations and generated explanations. However, these interpretations were necessarily cautious, because very few studies directly operationalized theory-linked conceptual mechanisms such as process internalization, object encapsulation, reification, or alignment between conceptual images and formal definitions. In LLM-supported contexts, gains in explanation quality coexisted with risks of procedural outsourcing when students relied on generated solutions without prior reasoning. By contrast, ITS-based environments more often supported tightly structured procedural engagement, suggesting that different AI modalities afford different forms of cognitive support and risk. Overall, AI’s conceptual impact appears to depend less on tool availability and more on instructional orchestration (task design, prompting, and teacher mediation). The findings also suggest that sustainability-related dimensions—particularly learner agency, transparency of AI support, and equitable participation—are closely connected to whether AI use promotes durable conceptual learning rather than superficial performance gains. Future research should operationalize cognitive transitions, assess structural understanding, and report AI-use conditions transparently to support cumulative, theory-driven synthesis.