Vibe coding with preservice teachers: a single-group pretest-posttest pilot study of programming self-efficacy and intentions to use generative AI in education

Abstract

Generative artificial intelligence (GAI) is changing how novices engage with programming by enabling program creation through natural-language interaction, an approach increasingly described as vibe coding. This may be particularly relevant for preservice teachers, who are often expected to develop computational skills despite limited programming experience. Grounded in social cognitive theory, this study examined changes in programming self-efficacy and intentions to use GAI in education following a four-week vibe coding intervention. A single-group pretest-posttest design was used with 159 preservice teachers (92.5% women) from four undergraduate programs at a Turkish state university. The intervention combined instructor demonstrations with individual programming tasks completed through natural-language prompting. Programming self-efficacy for simple and complex tasks and, as a secondary outcome, intentions to use GAI in education were assessed before and after the intervention. Because full measurement invariance was not established, pretest-posttest differences were interpreted as changes in observed scores. Significant increases were found for simple-task self-efficacy (Cohen’s dz = 2.62), complex-task self-efficacy (Cohen’s dz = 1.72), and intentions to use GAI in education (Cohen’s dz = 0.29), with all comparisons remaining significant after Bonferroni correction. The findings provide preliminary evidence that structured vibe coding activities may be associated with greater programming-related confidence among preservice teachers. However, the absence of a control group prevents attribution of the observed changes to the intervention, and the large self-efficacy effects may partly reflect very low pretest scores and regression to the mean. In addition, the CPSES measures general programming confidence rather than unaided programming ability, so posttest gains may partly reflect confidence in completing tasks with GAI support. Future studies should use controlled designs and performance-based measures to determine whether increased confidence is accompanied by gains in independent programming competence.