Lost in Code Generation: Reimagining the Role of Software Models in AI-driven Software Engineering

TL;DR

The paper investigates vibe coding, where natural-language prompts and AI coding assistants enable rapid creation of software, highlighting benefits in accessibility but exposing deep risks to robustness, security, and maintainability. It proposes reimagining software models as post-hoc, bidirectionally usable abstractions recovered from AI-generated code to restore understanding, expose risks, and guide refinement. The authors outline a framework for bidirectional co-evolution of code and abstractions, ensuring models mediate human intent and long-term evolution, and discuss agentic workflows where AI systems actively monitor and adapt software. A research agenda addresses fidelity, abstraction levels, trust, verification, and accessibility to non-experts, aiming to enable sustainable AI-driven engineering rather than one-off code generation.

Abstract

Generative AI enables rapid ``vibe coding," where natural language prompts yield working software systems. While this lowers barriers to software creation, it also collapses the boundary between prototypes and engineered software, leading to fragile systems that lack robustness, security, and maintainability. We argue that this shift motivates a reimagining of software models. Rather than serving only as upfront blueprints, models can be recovered post-hoc from AI-generated code to restore comprehension, expose risks, and guide refinement. In this role, models serve as mediators between human intent, AI generation, and long-term system evolution, providing a path toward sustainable AI-driven software engineering.