Human-AI interactions (HAII) in STEM education: A systematic review from 2010 to 2025

Abstract

Artificial intelligence (AI) is increasingly integrated into STEM education to construct adaptive learning environments. The perspective of Human-AI Interaction (HAII) offers a critical lens to examine how agency and roles are distributed between students and AI systems, highlighting the alignment of technological designs with pedagogical goals. To clarify these dynamics, this systematic review followed the PRISMA guidelines to select 327 empirical studies related to HAII in STEM education published between 2010 and 2025. Specifically, this review proposed a conceptual framework centered on HAII relationship paradigms, reviewing the educational contexts, technical affordances, interaction types & modalities, and learning outcomes of each. First, in the AI-directed, learner-as-recipient paradigm, AI systems exercise strong control over student agency. It retained a relatively higher presence in primary school and mathematics contexts, relying on probability-based models and compulsory designs to direct learning pathways. Employing implicit data tracking and interface controls, the evaluations primarily focused on domain knowledge and system usability. Second, in the AI-supported, learner-as-collaborator paradigm, agency is shared between students and AI systems. With relatively increased implementation in secondary education and science, these systems employed a mix of algorithms to provide adaptive scaffolding. Combining explicit text inputs with sustained implicit tracking, this paradigm broadened assessments to include affective outcomes and general skills. Third, in the AI-empowered, learner-as-leader paradigm, students hold the highest autonomy. Heavily concentrated in higher education, these systems leveraged Generative AI and suggestive designs to support explicit on-demand inquiry through natural language. The reliance on implicit tracking declined significantly, and the assessment of traditional domain knowledge decreased relative to other paradigms. Based on these results, this review offered several implications to inform the future design of AI applications in STEM education.