The Prisoner's Dilemma of AI: Cognitive Offloading Among Future Mathematics Teachers

Abstract

The integration of Large Language Models (LLMs) into education has introduced substantial opportunities as well as challenges related to students’ cognitive processes and professional development. For pre-service teachers, generative Artificial Intelligence (AI) creates a professional paradox. While AI systems can increase efficiency, support learning processes, and provide rapid access to information, excessive reliance on such systems may lead to cognitive offloading, automation bias, and gradual weakening of critical thinking and professional autonomy. Within educational contexts, this tension may be interpreted as a conflict between short-term individual gains and long-term consequences for human expertise and independent reasoning. This study compares pretest and posttest perceptions of generative AI among first-year pre-service mathematics teachers following a targeted educational intervention focused on principles of AI functioning and related cognitive risks. The Prisoner's Dilemma is used as a conceptual and pedagogical lens for framing the tension between the immediate benefits of AI use and its possible long-term educational consequences; it is not empirically operationalised or tested as a game-theoretic model. Within this framework, delegating cognitive activities to AI may represent a rational strategy at the individual level while simultaneously producing undesirable outcomes for broader educational development. A quasi-experimental pretest–posttest design without a control group was employed. Participants consisted of undergraduate students enrolled in a Mathematics Teacher Education programme at the University of Ostrava (pretest n = 18; posttest n = 14). Data collection was conducted using a semantic differential instrument consisting of sixteen bipolar dimensions measuring students’ perceptions of AI characteristics. The intervention included instructional activities explaining principles of large language models, cognitive offloading, automation bias, and ethical implications of excessive dependence on AI systems. Data were analysed using descriptive statistics and non-parametric methods, specifically the Mann–Whitney U test. The results did not reveal statistically significant differences between pretest and posttest measurements across any of the analysed dimensions. Several mean scores differed descriptively between the measurements, particularly for transparency, predictability, and the role of AI in deeper learning, but these differences were not statistically significant. Given the small, unmatched samples and the absence of a control group, they cannot be interpreted as evidence of an intervention effect. These results highlight the importance of integrating critical AI literacy into teacher education programmes.