GENERATIVE AI REDUCES PERIPHERAL BUT NOT MENTAL COGNITIVE LOAD IN PROGRAMMING EDUCATION

Abstract

Generative AI tools have entered programming classrooms faster than universities have learned to manage them, and evidence from resource-constrained education systems remains scarce. This study examined how engineering students use ChatGPT when learning programming and whether this use is associated with changes in perceived cognitive load. A quantitative cross-sectional survey design was applied: 140 undergraduate engineering students completed a three-block questionnaire combining a demographic and usage profile, a modified NASA-TLX cognitive load scale, and a perceived usefulness scale adapted from the Technology Acceptance Model. The results show near-universal adoption: 94.3% of respondents use ChatGPT for learning, mainly for code debugging and explanations of theoretical concepts. Compared with conventional study, working with ChatGPT was rated significantly lower on five of six NASA-TLX subscales (physical demand, temporal demand, frustration, perceived difficulty, and effort), whereas mental demand did not differ. Usefulness ratings were moderate, risk awareness was low, and usage intensity declined from the second to the fourth year of study (H=7.78; p=0.021). The novelty of this study lies in providing the first field evidence from Central Asia that self-directed generative AI use selectively removes the peripheral load of programming work while leaving its intellectual core intact. The main limitations are the single-institution sample and the use of self-reported measures. The findings support scaffolded integration of AI tutors and compulsory modules on verifying AI-generated code within AI literacy provision.