A Large-Scale Empirical Study of AI-Generated Code in Real-World Repositories

9 hours ago

The study conducts a large-scale measurement comparing AI-generated and human-written code in real-world repositories.
It addresses limitations of prior lab-based studies with synthetic benchmarks and small-scale tasks.
Metrics cover code-level aspects like complexity, style, security, and commit-level traits such as size, frequency, and stability.
Findings reveal small differences in code-level metrics between AI and human code in real-world settings.
Insights include variance in security quality across programming languages and new data on code duplication and commit characteristics.
The results provide implications for AI-assisted programming practices.

Hasty Briefsbeta