AI-assisted software development: Techniques, impact, challenges, and future directions

Abstract

Software development practice is being transformed by large language models (LLMs), yet the field lacks consolidated guidance on how, and under what conditions, these tools actually deliver reliable results. This paper reports a semi-systematic synthesis of AI-in-software-engineering research spanning 2018–2026, complemented by a hands-on prompting experiment. We pitted one-shot and iterative prompting against each other across three back-end coding tasks using a GPT-4-class assistant. The iterative approach cut defects by 83% and reached complete functional correctness on two of the three tasks, while one-shot prompting failed to achieve full correctness on any. That improvement, however, came with a time cost of roughly three to four extra minutes per task. These numbers matter less as benchmarks than as evidence of a trade-off practitioners need to consciously manage. The literature confirms the other side of the coin: AI assistance introduces hallucination, automation bias, and a subtler long-term risk of skill erosion that standard testing pipelines do not catch. Drawing on both the experimental data and the synthesis, we introduce the RAISE Framework (Responsible AI Integration in Software Engineering), a four-layer model—spanning AI capability assessment, SDLC integration, governance, and sociotechnical factors—intended to give teams a structured way to capture the benefits while keeping the risks in check.