AI Can Now Write Reusable Rules to Fix Code Automatically

Have you ever waited for an app update because a bug was taking forever to fix? Behind the scenes, developers often spend days or weeks manually scanning thousands of code files to fix problems, upgrade software versions, or repair broken features. AI is already used to suggest one-off fixes, but that gets slow and expensive when done file by file.

This new study tested whether AI could do something smarter: learn from a few examples and write a general rule — think of it as a smart find-and-replace instruction — that can automatically apply the same fix to every relevant file in the codebase. The researchers tried three AI models, including the powerful GPT-5.4 and smaller, more affordable open-weight models like Llama 3.1. Their biggest test: fixing API misuse, migrating old code to new language versions, and repairing programs.

The results are promising. GPT-5.4 wrote rules that were applied successfully across many different coding tasks and came closest to what human experts would have written. Smaller AI models did fine on simple, localized tweaks but got tangled in complex, cross-file migrations. This shows that AI isn't just good at doing one task at a time — it can understand a problem well enough to create a reusable playbook that saves developers from repeating the same work over and over.

Why should you care? Software runs your bank, your car, your workplace tools. When companies can update systems faster and cheaper, you see fewer outages, quicker bug fixes, and lower costs that are often passed on to consumers. The catch: AI rules aren't perfect yet. They occasionally miss edge cases, and big models are costly to run. Still, this research shows we're moving closer to a future where AI pulls most of the weight in routine software maintenance, letting humans handle the tricky judgment calls.