Exploring Impact of ChatGPT in Software Development : An Experimental Study on Code Generation and Legacy Code Improvement
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
This thesis explores how far ChatGPT (GPT-4) can be used as a practical software-development assistant from an end-user perspective, and whether it can perform key development tasks autonomously.The study focuses on two common scenarios: generating a small web application from prompts and refactoring a legacy codebase to modern standards.Two experiments were conducted using ChatGPT Plus (GPT-4) in a controlled setting (no internet access or plugins) to keep conditions consistent.In the first experiment, GPT-4 was prompted to produce a simple React-based weather application.The generated project compiled and displayed the intended user interface but initially failed to show weather results due to an incorrect API endpoint, requiring manual debugging and correction.In the second experiment, GPT-4 was asked to refactor a legacy JavaScript browser game.The refactored code produced a runnable structure and displayed the game frame, yet core gameplay behavior such as falling pieces and user controls remained incomplete or incorrect.Attempts to repair these issues through iterative prompting improved some parts but also introduced inconsistent or repetitive codes, showing the limits of refactoring complex behavior without deeper context.Overall, the results indicate that GPT-4 can meaningfully speed up early development by generating structure, boilerplate, and partial implementations, but it does not yet deliver complete reliable functional correctness without human oversight.The experiments highlight the importance of well-structured prompting and reinforce the need for a human-in-the-loop workflow in which developers validate outputs, debug faults, and supply missing context.The thesis concludes that GPT-4 is currently best understood as an augmentation tool that supports developer productivity rather than a replacement for human software engineering, and it discusses implications for practice and education, particularly around prompt literacy, verification, and responsible use.
Fields of Study & Subjects
Software Engineering Research
Software Engineering Techniques and Practices
Artificial Intelligence Applications
Publication details
- Journal / source
- Tampere University Institutional Repository (Tampere University)
- Publication date
- 1 January 2026
- Language
- EN
- Document type
- Thesis