AI-assisted code generation tools: a new frontier in software development

Abstract

Through various developments in Artificial Intelligence code generation, it has greatly influenced the development of codebases and software. This has increased efficiency and led to the development of conventional programs and practices. This thesis emphasises these AI tools and their impact on the developer’s experience in real-world applications, and also addresses critical gaps in the existing literature. Through thorough analysis, this paper demonstrates that AI code generation can enhance learning, efficiency and job satisfaction among young developers. However, studies show that these AI-generated codes may be inaccurate at times and also contain code smells, which can turn into a hassle for developers to debug rather than a blessing. Therefore, a more accessible pricing strategy and improved reliability should be implemented in these tools. Our research includes three distinct pipelines that help automate code debugging, followed by automated unit tests after the code has been fixed. These pipelines include Prompt Mechanism, which uses Abstract Syntax Trees (AST) and Pylint for detecting errors and with the help of a generative AI it fixes the code accordingly to the detected error, a Retrieval-Augmented Generation (RAG) pipeline leveraging official Python documentation for context-aware corrections and a Fine-tuning approach using CodeT5 trained on samples from the TSSB-3M dataset. Afterwards, the accuracies of these pipelines were evaluated by examining syntax, runtime, and logical errors. The fine-tuning and RAG pipeline had similar accuracy to one another, while the prompt mechanism had lower accuracy. By analysing code-practising factors such as usefulness, efficiency, and adaptability across various tools and techniques used in Artificial Intelligence, the study examines both the advantages and negative aspects of using AI tools in large-scale industrial development. While AI tools can be an excellent choice for standard coding tasks, human software quality assurance is also necessary to maintain the accuracy of the generated code. This thesis research helps to make use of responsible AI practices and addresses the concerns regarding the security and the quality of the codes, and also makes sure to include the ethical implications through a clear explanation, which would help the developers to learn and unit tests, which would require less human oversight for quality assurance. This thesis demonstrates that a multiple pipeline approach can be beneficial in reducing errors, increasing accuracy, and enhancing robustness in a developer’s code.