Generative Ai In Software Engineering: A Meta-Analysis Of Developer Productivity, Code Quality And Verification Challenges

Abstract

The rapid integration of artificial intelligence (AI) into software engineering workflows, particularly through AI 
pair-programming assistants and code-generation copilots, has prompted a pressing need to empirically evaluate 
whether such tools genuinely improve development outcomes relative to conventional, manually driven practices. 
This study presents a controlled empirical investigation comparing AI-assisted software development with 
conventional development across five real-world project categories, involving forty professional developers 
organized into ten matched teams over a six-sprint development cycle. Quantitative data were collected on task 
completion time, defect density, code maintainability, test coverage, developer productivity, and developer 
satisfaction, and were analyzed using descriptive statistics, paired t-tests, and correlation analysis. The results 
indicate that AI-assisted teams completed tasks 32.6% faster on average, exhibited a 38.4% lower defect density 
by the final sprint, and reported higher satisfaction scores across ease of use, code confidence, and learning curve 
dimensions, while conventional teams retained a modest advantage in architectural design depth for highly novel 
problem domains. Statistical tests confirmed that the observed differences in completion time and defect density 
were significant (p < 0.05), supporting the hypothesis that AI assistance meaningfully accelerates delivery without 
compromising, and in several respects improving, code quality. These findings extend prior comparative studies 
by triangulating five distinct empirical measures within a single controlled design and offer practical guidance 
for organizations considering the adoption of AI-assisted development pipelines, while also highlighting boundary 
conditions under which conventional expertise remains indispensable.