Beyond Functional Correctness: Design Issues in AI IDE-Generated Large-Scale Projects

A new study led by researchers from multiple universities provides the first empirical evidence on the project-level capabilities and pitfalls of AI-powered IDEs like Cursor. Using a novel Feature-Driven Human-In-The-Loop (FD-HITL) framework to systematically guide generation, the team created 10 large-scale projects across three domains. The results were a mixed bag: while the AI demonstrated impressive scale, producing projects averaging 16,965 lines of code and 114 files, and achieving a high 91% average functional correctness score, this came at a significant design cost.

A deep dive into the generated code using static analysis tools CodeScene and SonarQube uncovered a staggering number of design flaws. The analysis identified 1,305 issues across 9 categories with CodeScene and 3,193 issues across 11 categories with SonarQube. The most prevalent problems were Code Duplication, High Code Complexity, Large Methods, and violations of framework best practices and accessibility standards. These issues systematically violate fundamental software design principles such as the Single Responsibility Principle (SRP), Separation of Concerns (SoC), and Don't Repeat Yourself (DRY).

The study concludes that while agentic AI IDEs are powerful tools for generating functional software at scale, they currently act as 'junior developers' who lack architectural foresight. The generated code works but is riddled with technical debt, creating substantial risks for long-term maintainability, evolvability, and scalability. The authors emphasize that successful integration requires careful review and refactoring by experienced human developers to mitigate these design risks before such projects can be considered production-ready.