Exploring the Efficacy of Large Language Models in Full-stack Development : A Study on the Quality of ChatGPT Generated Code in Full-stack Development

Sammanfattning

Background The rapid advancement of artificial intelligence (AI) and machine learning (ML) has led to tools like ChatGPT becoming increasingly common in professional software development. As adoption grows, so do questions about the quality and reliability of the code these tools produce. Problem Despite the growing interest in using AI tools in development workflows, there is limited empirical evidence on how LLM-generated code holds up when evaluated against professional standards. The quality, security and maintainability of AI-generated code in full-stack development is still insufficiently understood. Objectives This study aims to evaluate the quality of code generated by ChatGPTin a full-stack development context through structured code reviews by experienced developers. The goal is to identify recurring strengths and weaknesses and derive practical recommendations for integrating ChatGPT into professional software development workflows. Research Method A full-stack web application was developed entirely through interactions with ChatGPT (GPT-3.5), without any manual coding. The codebase was then reviewed by five professional software developers who evaluated it freelybased on their own professional judgment. The feedback was analyzed using thematic analysis to identify recurring patterns. Results The reviews show that ChatGPT produces readable, reasonably structured code with some good practices in place. Password hashing, route protection and clean a sync handling were all noted as strengths. However, significant issues were also identified: inconsistent documentation, missing infrastructure, no input validation, and security vulnerabilities in the authentication flow. The most critical finding was that the login functionality does not work at all, due to mismatches between the server and client code caused by the model losing contextual awareness over the course of a long conversation. The study concludes that ChatGPT works best as a first-draft tool, and that structured human review and end-to-end testing are essential before AI-generated code can be trusted in a production settings.