Generative AI Assistants in Software Development: Productivity Gains, Code Quality, and Security Trade-offs

Abstract

Generative artificial intelligence (AI) coding assistants promise to shorten software-development tasks, yet task speed alone cannot establish an improvement in engineering productivity. This critical narrative review asks whether assistants help developers complete work faster without increasing functional defects, maintenance burden, or security weaknesses. The final analytical corpus comprises 20 principal sources available through September 27, 2026, including experiments, field trials, repository studies, a benchmark, methodological work, and security guidance. Findings remain heterogeneous: a bounded JavaScript task was completed 55.8% faster with Copilot; pooled workplace trials found 26.08% more completed tasks; an observational Phase 1 maintainability result showed a 30.7% shorter median initial-task time; and experienced maintainers in mature repositories took 19% longer when AI was permitted. A 2026 controlled interface study found shorter task times while directly recording verification-related interactions. A later METR update suggests tool-generation change but cannot yield a precise productivity estimate because of selection and measurement problems. We propose Quality-Adjusted Development Productivity (QADP) as an author-developed decision framework for measuring the complete task lifecycle under independent quality gates, not as a validated score. The evidence supports conditional gains, with task, developer, interface, verification burden, and downstream effects defining the applicable boundary.