Navigating AI in STEM: What Secondary Students Actually Do With Generative AI-Driven Tools

Synthesis: This cross-sectional, convergent mixed-methods study (n = 416 Czech secondary students) maps what students actually do with generative AI tools across four STEM subjects, revealing a stratified adoption pattern: applied disciplines (computer science, economics) normalize AI as a collaborative resource, while theoretically rigorous subjects (mathematics, natural sciences) show high perceived prohibition co-occurring with persistent clandestine use and poor rule clarity. Students mostly position AI as an instrumental scaffold for explanation and verification — but a critical evaluation gap emerges: heavy operational prompt modification overshadows external factual verification, shifting behavior toward rapid cognitive offloading.

Key Findings

1. A stratified, divergent adoption model across subjects. Computer science (51.2% see AI as an integrated legitimate tool; only 8.2% prohibit) and economics normalize AI use, while mathematics (65.9%) and basic natural sciences (55.3%) report high prohibition (Kruskal-Wallis H = 360.15, p < .001). Institutional AI AI Governance reflects localized disciplinary cultures rather than a unified policy.
2. A transparency gap in restrictive subjects. Mathematics and natural sciences are perceived as highly restrictive but poorly defined (37.3% and 40.9% report complete lack of clarity); rule clarity is significantly discipline-dependent (χ² = 101.65, df = 12, p < .001). Restrictive subjects combine de facto prohibition with unclear guidelines, fueling uncertainty and inequities from implicit norms.
3. AI is predominantly an instrumental scaffold, not a substitute. Most students use AI for step-by-step theory explanations (n = 265), checking their own solutions (n = 225), and task/solution analysis (n = 202) — rarely as a full substitute (significant replacement < 10% across cohorts). ChatGPT dominates (n = 364), near-universal regardless of cohort.
4. A behavioral asymmetry: operational prompting overshadows epistemic verification. Though students frequently encounter incorrect/hallucinated outputs (especially in mathematics, n = 170), systematic cross-verification with external sources is reported at low frequencies. Prompt refinement dominates, so the operational phase of the workflow overshadows the epistemic verification phase — shifting behavior toward rapid cognitive offloading.
5. Evolutionary, not disruptive, integration. A cross-cohort shift from task substitution (first-year students) toward strategic augmentation and advisory/verification modes (third/fourth years) indicates AI functions as a cognitive mediator that accelerates workflow while fostering "operational dependency."

What this means for practice

• Learners. Verify AI output against an external source as a routine step, not an occasional one: the study found continuous prompt modification heavily overshadowed external factual validation, and systematic cross-checking was reported at low frequencies even though students regularly met algorithmic errors and procedural dissonance.
• Learners. Keep treating AI as an instrumental scaffold for explanation and procedural verification rather than a substitute for independent reasoning — the position most students reported — because the risk the study surfaces is operational dependency that accelerates the workflow while thinning the discipline.
• Instructors. Make the subject's rules explicit and subject-sensitive, because mathematics and natural sciences combined high perceived prohibition with unclear guidance, so implicit norms drove clandestine use and inequity rather than honest disclosure.
• Administrators. Move from restrictive governance to a framework that formally integrates operational and epistemic workflow competencies, since the authors argue bans face reverse-causality reporting biases, and fund AI Literacy instruction that strengthens cross-verification and evaluation skills rather than tool access alone.

Limitations

• The study is cross-sectional and rests on self-reported behavior from n = 416 students at business-oriented secondary schools in the Czech Republic, so the patterns are associational and cannot support causal claims; the authors note that restrictive governance carries distinct reverse-causality reporting biases.
• Adoption is measured at the level of four STEM subjects used as proxies for disciplinary epistemologies and task environments, so subject differences may reflect school- or cohort-level conditions rather than the disciplines themselves.
• Clandestine use and rule clarity are self-reported in settings where students perceive AI use as prohibited, so those frequencies may be misreported in either direction.
• The survey and thematic reflections come from one national context and one school type, leaving transfer to other educational systems untested.

Connected Concepts

• Generative AI
• K-12
• STEM Education
• AI Literacy
• Student Experience
• Cognitive Offloading
• Student-AI Interaction
• Prompt Engineering
• Scaffolding
• Educational AI Policy
• AI Governance
• Equity

Connected Articles

• Young People, Learning, and Generative AI: A Rapid Literature Review and Implications for PreK-12 Education — GenAI across learners, contexts, cultures
• The Generative AI Learning Penalty: Evidence from Chinese Secondary Education — Generative AI learning penalty in secondary students
• Generative AI in Higher Education: A Systematic Review of Opportunities, Challenges, and Pedagogical Innovations (2022–2025) — GenAI in higher education systematic review
• Conversational AI agents in education: an umbrella review of current utilization, challenges, and future directions — Umbrella review of conversational AI agents in education

Citation

Lnenicka, M., & Coufal, P. (2026). Navigating AI in STEM: what secondary students actually do with generative AI-driven tools. International Journal of STEM Education, 13, 36.