Vibecoding
New Research Reveals Vibe Coding Is Iterative, Not One-Shot—With Mixed Evidence on Long-Term Quality
A comprehensive review of 47 academic and industry sources shows developers using AI-driven code generation spend more time supervising and validating than writing, but concerns about maintainability and safety remain largely unanswered.
Vibe coding—the practice of describing software intent in natural language and letting large language models generate the code—is far more nuanced than the popular notion of a single AI prompt. A new multivocal literature review spanning academic papers and industry sources reveals the reality: developers engage in a continuous loop of generating, evaluating, and refining AI-produced code, fundamentally shifting how they work.
What the Review Found#
Researchers analyzed 47 sources (28 peer-reviewed, 19 from industry) published between 2022 and October 2025. The study paints a consistent picture: vibe coding is an iterative workflow, not a one-shot transaction. Developer time increasingly goes toward specification, validation, and oversight rather than writing code from scratch.
On the productivity side, 21 of 47 sources (45%) reported short-term gains in velocity and time-to-prototype. However, the evidence becomes sparse when examining what happens next—maintainability, long-term code quality, and the effectiveness of safeguards remain largely unexplored.
Where Vibe Coding Works—and Where It Doesn't#
The practice shows the strongest track record in prototyping and user interface development, domains where iteration and rapid feedback loops are already the norm. In contrast, adoption in production systems, data-intensive pipelines, and safety-critical applications remains weak, suggesting developers and organizations are appropriately cautious about AI-generated code in high-stakes contexts.
The review also found that tool visibility does not guarantee effectiveness—a reminder that popularity and actual quality are separate concerns.
What's Still Unknown#
Major research gaps persist. Questions about how developers actually govern and oversee generated code during sessions, the long-term maintenance burden of AI-written systems, and whether built-in safeguards genuinely prevent defects remain largely unanswered. The authors call for future studies to address these gaps, particularly in production, data-heavy, and safety-critical domains where the stakes are highest.
