The emergence of ‘vibe coding’ is reshaping the landscape of software development, allowing individuals without traditional programming skills to create applications through artificial intelligence (AI) prompts. However, research from ETH Zurich indicates that merely relying on AI for inspiration is insufficient for successful software creation.
‘Vibe coding’ enables users to describe the functionality of a program in natural language, while AI tools like Claude Code, Cursor, and Loveable generate the corresponding code in the background. This shift marks a significant departure from the past, when programming was largely restricted to those proficient in languages such as Python or Java.
Researchers at ETH Zurich identified two critical factors influencing success in vibe coding: foundational knowledge of computer science and clarity of expression. Understanding the basic principles of how programs operate is essential, as is the ability to articulate ideas clearly in writing.
In an experiment involving 100 students tasked with recreating existing applications using AI agents, findings revealed that knowledge of computer science had the most significant impact on the quality of the final product. The study’s authors noted that informed users are better at planning application structures and identifying errors quickly.
“Experts use precise technical terminology, allowing the AI agent to operate more effectively, even when the user does not see the underlying code,” explained doctoral student Theo Weidmann.
Interestingly, the study also found that students who frequently relied on neural networks in their daily lives performed worse on the tasks. Researchers suggest that over-reliance on AI may weaken an individual’s ability to articulate their thoughts clearly. In vibe coding, where prompts effectively become a form of code, imprecise wording can lead to software errors.
Another study led by Professor Martin Vechev highlighted a significant issue regarding the efficacy of AI agents in debugging code. The results were striking: in over 70% of cases, AI attempted to ‘fix’ code that was already correct. Vechev concluded that this tendency creates a risk of introducing new errors into functioning systems.
“This serves as a reminder that human experts must continue to review AI-generated code rather than relying on it blindly,” noted the researcher. Although vibe coding lowers the barrier to entry in the IT field, the full automation of software development remains a future consideration. Experts believe that the role of humans as overseers and architects of AI will continue to be paramount.
The rise of 'vibe coding' allows non-programmers to develop applications using AI, but research shows that foundational computer science knowledge and clear communication are crucial for success. Furthermore, reliance on AI can lead to errors, highlighting the ongoing need for human oversight in software development.