Refatoração de código: comparação entre abordagens manuais e inteligência artificial generativa

Abstract

Code refactoring is an essential practice in Software Engineering, aimed at improving the internal structure of code without altering its external behavior, thereby contributing to the quality, maintainability, and long-term evolution of software systems. With the advancement of Generative Artificial Intelligence tools, the possibility of automating this process has emerged, raising questions about the effectiveness and limitations of such approaches in comparison to human intervention. This paper presents a comparison between manual code refactoring and refactoring assisted by GitHub Copilot and Amazon Q Developer, using three educational games developed within the Educational Software Development Project at the Federal University of Technology – Paraná (UTFPR) as case studies. Manual refactoring was conducted based on classical techniques described by Fowler, while automated refactoring was performed using a standardized prompt applied to both tools. The results were evaluated through software quality metrics obtained using SonarQube, considering attributes related to maintainability, comprehensibility, reusability, and structural complexity. The findings indicate that Generative AI tools were capable of achieving significant improvements in several quality metrics. However, they also exhibited limitations related to context preservation, generation of compilation errors, and the introduction of modifications that extend beyond the traditional scope of refactoring. Furthermore, each tool demonstrated superior performance in specific quality attributes, whereas the manual approach provided greater control over architectural decisions and the targeted application of refactoring techniques. The results contribute to a better understanding of the potential and limitations of Generative Artificial Intelligence as a support mechanism for software maintenance and evolution.