AI coding transforms software development, insights from 57 industry worker videos

The rapid evolution of software development, driven by large-scale language models, presents both opportunities and challenges for computing professionals and educators. New research investigates how these tools are reshaping industrial practices. Hung-Fu Chang from the University of Indianapolis, Mohammad Shokrolah Srizi and Lizhou Cao from the University of Maryland Eastern Shore, and Supannika Koolmanojwong Mobasser from the Boehm Center for Systems Software Engineering analyzed input from software practitioners shared in online videos to understand current usage, associated risks, and changes to established workflows. While the study found that these new coding techniques significantly increased productivity and reduced barriers to entry, it also highlighted new concerns about code quality, security, and the potential for compromising basic programming skills. This industry-focused analysis provides important insights for adapting computer science and software engineering curricula to prepare future engineers for rapidly changing professional environments, advocating greater emphasis on problem solving, architectural design, and rigorous code review practices.

Vibe coding, themes and critical perspectives

A comprehensive analysis of YouTube videos reveals different perspectives on Vibe coding, a rapidly evolving approach to software development. These videos present a wide range of opinions that fall broadly into Vibe coding pro-, critical, or nuanced categories, and include additional content that focuses on explanations, the future of the practice, and security considerations. The videos focus on core themes such as speed, maintainability, etc. While many practitioners emphasize the potential benefits of Vibe coding, others have expressed concerns about its long-term viability and potential technical debt. A balanced perspective recognizes both benefits and risks and focuses on best practices for safe and effective implementation. A significant number of videos explore the future of Vibe coding as it relates to artificial intelligence, agents, and broader changes in software development. These discussions often center on AI's potential to automate coding tasks, lower barriers to entry, and accelerate innovation, while also raising questions about the evolving role of developers and the need for new skills.

YouTube videos that reveal AI coding practices

Researchers conducted a new study on AI-assisted coding by analyzing content shared by professional developers on YouTube. They have curated a collection of 57 videos published between late 2024 and 2025 that focus on hands-on experience with large-scale language models (LLM) in real-world software development. This approach addresses a gap in existing research, which often focuses on educational applications of LLM coding tools. The team employed a rigorous qualitative analysis process to carefully select videos that accurately represent current industry practices. This in-depth study reveals the definition of both “vibe coding” and “agent coding” practices, distinguishing them from traditional code generation techniques.

According to industry reports, a significant portion of code in some organizations is generated using these technologies, demonstrating significant productivity gains and lower barriers to entry. This analysis highlights the shift in development bottlenecks to code reviews and concerns around code quality, security, and the potential for compromising basic programming skills. This study provides valuable insights into AI-assisted coding and provides guidance for adapting computer science and software engineering education to the demands of a rapidly changing professional environment.

LLM reimagines professional software development workflows

This study details how large-scale language models (LLM) are impacting professional software development, based on a qualitative study of 57 videos featuring experienced practitioners. This research marks a shift towards AI-assisted coding approaches such as vibe coding, AI-assisted coding, and agent coding, all of which are accelerating code generation and changing development challenges. The study found that development bottlenecks are shifting from early code creation to areas such as code review, testing, and ensuring system-level security. This highlights the growing importance of skills focused on evaluating and validating AI-generated code.

Practitioners describe “vibe coding” as emphasizing orchestration over direct coding and focusing on high-level instructions and functional output. “Agent Coding” delegates coding tasks to autonomous AI agents, simulating collaboration with multiple engineers. While AI tools accelerate prototyping and experimentation, they are raising concerns about code quality, maintainability, security vulnerabilities, and the potential for compromising basic problem-solving skills.

AI will reshape software development practices

This study presents a systematic analysis of how large-scale language models are reshaping software development, based on an in-depth examination of industry participants' experiences shared in online videos. This study identifies a shift to different coding approaches that support AI, demonstrating the acceleration of code generation and the associated changes in development challenges. Bottlenecks are shifting from initial code creation to areas such as code review, testing, and ensuring system-level security, highlighting the growing importance of skills focused on evaluating and validating AI-generated code. The findings also highlight changes in developer roles and skill requirements, with AI tools lowering barriers to entry but also raising concerns about potential skills decline. As a result, this study suggests a need for adaptation in computer science education, advocating increased use of project-based learning that incorporates AI tools and real-world workflows, as well as a greater emphasis on problem solving, architectural thinking, and specification-driven development. This study represents an early snapshot of a rapidly evolving field, and future research should focus on longitudinal studies to better understand the long-term co-evolution of AI-based development practices.