AI-Assisted Software Development: A Systematic Survey of Code Generation, Vibe Coding, and Human–AI Collaboration in Programming

The practice of programming is undergoing its most significant transformation since the advent of high-level languages. Large language models trained on code now write, complete, explain, and repair software inside the world's most popular development environments, and a new paradigm—popularly called vibe coding—lets people express intent in natural language while the model handles implementation. This paper presents a systematic survey of AI-assisted software development. We trace the evolution from statistical autocomplete to generative models, formalize the emerging interaction paradigms, and propose a two-dimensional taxonomy that separates tools by interaction paradigm (inline completion, conversational assistance, agentic execution) and by autonomy level (assistive, supervised, autonomous). We review representative models, benchmarks, and human-centered studies; examine how the community evaluates functional correctness, productivity, and security; and organize the risks—vulnerable generated code, testing debt, over-reliance, and licensing ambiguity—into a coherent framework. Finally, we identify open challenges, including reliable evaluation of agentic workflows, secure-by-default generation, and the changing skill profile of the software engineer, and outline directions for future research.