Artificial Intelligence in Physics Education: A Critical Integrative Review
Corresponding email: [email protected]
A B S T R A C T
Artificial intelligence (AI) is now used in physics education for tutoring, feedback, assessment, problem-solving, laboratory analysis, and teacher support, yet the literature often treats model capability, user satisfaction, and learning as equivalent. This review examines why similar technologies can lead to different educational consequences. A critical integrative design synthesized literature published from January 2018 to April 2026, while retaining earlier methodological sources where necessary. Physics-specific studies were prioritised, and science or STEM research was included only when it clarified mechanisms directly relevant to physics. A retrospective PRISMA-style audit identified 48 papers at the initial selection stage, of which 23 were included in the substantive critical synthesis. Six recurring functions were identified, but none were beneficial by definition. Positive effects were most credible when AI supported intermediate reasoning, operated within bounded tasks, drew on domain-relevant grounding, and remained subject to human verification. Evidence for affective and efficiency gains was more consistent than evidence for durable conceptual learning. The literature also revealed category errors in treating accurate answers as learning, conversational responsiveness as personalization, and fluent language as representational competence. Graphical, spatial, and multimodal tasks were especially vulnerable to reliability mismatch. The review proposes the Pedagogical Contingency Framework for AI in Physics Education, which explains outcomes through the interaction of pedagogical role, epistemic position, representational demand, and governance or grounding. Its central contribution is a move away from tool cataloguing towards an explanatory account of when AI augments disciplinary agency and when it substitutes for it.