Vibe Coding, Gatekeeping, and Learning to Program with ChatGPT

• In Event: Can Education Stop Killer Robots? Exploring Values, Ethics, and Impacts in K12 Computer Science Instruction

Abstract

Objectives
The allure of vibe coding – coding with an LLM in plain English without writing program code directly – is clear to students. IBM (2025), for example, has argued that it represents a way to democratize coding, opening computer code to those without access to the resources – teaching, social, and otherwise. Precursors to vibe coding (e.g., "plain language coding") have been a popular way to teach beginners computer programming since at least the 1970s (Biancuzzi, 2009).
In our study, high school students without coding experience vibe-coded to create playable games. We evaluated students' coding goals, their nascent CS understanding, how they used AI assistance, if/where they got frustrated in using the tools, and how they responded to that frustration. One goal in the study was to better understand the role that vibe coding might play in democratizing CS and/or AI.

Perspectives
This work is informed by Schwartz & Martin (2004) Preparation for Future Learning framework in which students engage in practices to find the limits of their own understanding. This project was designed as a constructionist design-based learning intervention – students iteratively built and refined a coding project to understand coding given specific constraints. This was part of a larger project exploring the sociocultural dimensions of youth interests of working, resisting, and learning alongside GenAI (Higgs & Stornaiuolo, 2024; McBride et al, 2024).

Methods
We met weekly with a small group of secondary students and a teacher in an afterschool setting and worked with them to use ChatGPT on student-led creative projects and games. The vibe-coding section started by asking students to use ChatGPT Canvas to code Chess and then either iterate on their own rules or build a game of their choice in ChatGPT Canvas. We collected both video and log data, including each session’s discussions between researchers, students, and the teacher. ChatGPT logs of student code and messages were saved by the team but not retained by OpenAI.

Results
Student reactions started as "excited" - the initial code generated by ChatGPT Canvas seemed to "do something" and seemed to give the students a springboard from which they could work. However, the code frequently failed, got convoluted, misunderstood student prompts, and "lost sight" of what had previously worked. In short, we found (in line with many other studies) that students lacked the computational literacy to work successfully in ChatGPT Canvas.

Scholarly significance
This is an early look at vibe coding as a method of CS instruction. AI is able to generate live, working apps in seconds such as games that students can actually play, but building something specific still seems to require more significant computational literacies. Preliminary analysis suggests that student perceptions of the difficulty of computer programming were changed by their experiences, with some participants perceiving it as "easier" and others "more complex" than before.

Authors

• Max Shafer-Landau, University of Wisconsin - Madison

• Matthew Berland, University of Wisconsin - Madison

• Daniel P. Moore, Seton Hall University

• Antero Garcia, Stanford University

• Nicole Mirra, University of California - Los Angeles