Year: 2026
Type: other
Abstract: The use of generative AI in software development has rapidly increased, and there are signs that AI-assisted software development might allow developers to produce software faster. In some applications, however, the software development process requires a non-author human developer to approve any suggested changes in a code review before the changes can be applied. Code reviews contribute to software quality in many ways, but they can also be time- and resource-consuming. This thesis examines how reliably an AI-generated T-SQL code review report highlights a defect or a potential query performance issue in a pull request. The purpose is to evaluate if generative AI can be considered as a reliable assistive tool that helps to identify issues that a human developer might have overlooked in a code review. The work was conducted in an enterprise-scale codebase by implementing four versions of an AI-driven code review tool using Visual Studio Code and the GitHub Copilot Chat extension. The reliability was evaluated by conducting three experiments with the four implemented versions, resulting in 120 T-SQL code review reports. The experiments were designed to replicate pull requests that were previously authored and reviewed by professional developers to evaluate how reliably an AI-generated code review report would have helped the reviewer to identify an issue within the pull request. The results suggest that an AI-generated code review report would have reliably helped the human reviewer to identify an issue which was overlooked in the original code review. The implemented code review versions also suggest that even a minimalistic code review agent skill may be beneficial in a small-scale code review, though including application context allows for a more in-depth review covering application-specific architecture and practices.
Source: Doria (University of Helsinki)
Author: Kauppinen, Tommy
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7160464678