Denis Tsyplakov of DataArt tested GitHub Copilot, JetBrains Junie, and Windsurf on a simple Java task. While AI tools excel at algorithmic code, they struggle with architectural decisions, producing overcomplicated, inefficient results when asked to implement thread-safe solutions or optimize for performance like using a TreeMap.

The experiment shows AI-generated code can become an unreadable mess if instructions are vague or prompts aren't fine-tuned. Developers must still possess strong code-reading skills to review every line, as these assistants aren't yet ready to replace human code review. The core issue is supportability—code that's hard to maintain slows down production updates.

For now, AI coding assistants work best with stable, structured languages like Java, C#, and TypeScript. They can speed up development by 10-20 times for specific tasks, but only when the developer already understands the desired outcome. Used blindly, they create technical debt; used thoughtfully, they augment skilled engineers.