Year: 2025
Type: dissertation
Abstract: This thesis explores how AI-generated code affects readability in software development.As LLMs become more common in the field, it is important to understand how readable the code they generate is compared to code written by humans.This study uses a comparative experimental approach by applying the same prompt to five models (ChatGPT-4, Claude-3.5,Microsoft Copilot, Gemini-2.0,and DeepSeek R1) across four programming languages: Java, JavaScript, TypeScript, and Python.Static code analysis tools such as SonarQube, PMD, ESLint, and Qlty were used to measure metrics such as complexity, clarity, structure, and comment quality.The results show that most AI-generated code is generally more readable than human-written code, although quality varies across models and programming languages.Claude-3.5 and ChatGPT-4 demonstrated the best overall performance, while Gemini-2.0showed weaker results.These findings provide useful insights for developers and teams selecting AI tools for code refactoring.
Authors: Derbas, Mazen, Samuel, Åhlander
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7000975733