Navigating AI in STEM: what secondary students actually do with generative AI-driven tools

The rapid proliferation of generative Artificial Intelligence (AI) into secondary education has outpaced formal institutional guidance, creating an expansive array of unsupervised, independent student digital practices. Moving beyond simplistic assumptions of technological affordances, this cross-sectional study empirically analyses the self-reported behavioral distributions of secondary students’ interaction with generative AI-driven tools across four STEM subjects: computer science, mathematics, natural sciences, and economics. Using a convergent parallel mixed-methods design, we synthesized survey data ( n = 416) and thematic reflections from business-oriented secondary schools in the Czech Republic to examine how institutional regulation and subject-specific environments shape student adoption. Quantitative findings reveal a stratified, divergent adoption model characterizing distinct academic fields when treated as structural and environmental proxies for varying disciplinary epistemologies and task environments. While applied and computational disciplines demonstrate normalized integration, theoretically rigorous subjects like mathematics exhibit a transparency gap characterized by high perceived prohibition and persistent clandestine use. We map the self-reported functional roles of AI in the student workflow, finding that students predominantly position AI as an instrumental scaffold for explanation and procedural verification rather than a direct substitute for independent reasoning. However, qualitative evidence uncovers a critical evaluation gap that exposes a behavioral asymmetry within the student workflow: operational prompting loops and epistemic verification are shown to function not as independent psychometric traits, but as interdependent workflow phases where continuous prompt modification heavily overshadows external factual validation. Although students frequently encounter algorithmic errors and procedural dissonance, where AI logic diverges from curricular standards, systematic cross-verification remains reported at low frequencies. The study suggests that AI integration is currently evolutionary rather than disruptive, often functioning as a cognitive mediator that accelerates workflow while shifting the behavioral layout toward operational dependency. These associational patterns highlight the need to transition from restrictive governance, which faces distinct reverse-causality reporting biases, to a framework that formally integrates overlapping operational and epistemic workflow competencies, emphasizing the need for subject-sensitive guidance that ensures AI enhances rather than displaces disciplinary rigor.