AI Can Now Draft Software Blueprints and Fix Its Own Errors

When you use an app, its software was designed using something like an architectural blueprint. For programmers, that blueprint is called a UML class diagram (a visual map of the software's structure). Normally, a human expert creates these diagrams, which takes time and skill. Now, researchers are testing whether AI can do this job—and how well it can fix its own mistakes.

The study, led by Jie Liang and Peng Liang, asked several popular AI models (the same kind of technology behind ChatGPT) to create UML class diagrams for different software systems. Then they compared AI's work to diagrams made by human experts. The results showed AI isn't ready to replace experts yet. It often missed important parts, drew confusing or missing relationships between components, forgot to include inheritance (a way to share code between related parts), and even created useless classes or circular dependencies (where parts rely on each other in a loop).

But the researchers didn't stop at pointing out problems. They designed three repair strategies, like giving AI a better memory, injecting outside knowledge, and automatically finding and fixing flaws. After applying these fixes, every AI model improved. On average, AI correctly identified 85% of key classes. More impressively, it fixed 100% of unused classes and circular dependencies, though relationship repair was lower at 46%—showing that some problems still need human judgment.

For everyday people, this means AI could soon help software teams build programs faster and with fewer bugs. It won't replace human software architects, but it could act like a smart assistant that drafts the blueprint and learns from its mistakes. That could lead to cheaper, higher-quality apps—and let human experts focus on the harder, creative decisions.