ChatGPT Can Fix Code Bugs — But Its 'Self-Correction' Trick Often Fails

Software bugs cost companies enormous amounts of money, so there's a booming market in AI tools that promise to find and fix them automatically. This study examined three well-known ways of squeezing better results out of ChatGPT without retraining it — mostly "self-correction" (the AI reviews and rewrites its own fix) and "agents" (AI that can run commands and test its own work). These methods have reported impressive results, and that's how they get sold to engineering teams.

The researchers ran all three methods against three different collections of real-world broken code, and the results were all over the map. The same method that produced big improvements on one collection delivered no benefit — sometimes a negative one — on another. When they swapped in a newer version of the AI model, the pattern flipped entirely: a method that struggled before suddenly did well, and vice versa. In other words, the headline numbers depended heavily on which test you happened to pick.

Then came the most awkward finding. On one collection of buggy code, simply giving ChatGPT the error message and the failing test results produced more correct fixes than any of the elaborate self-improvement techniques. The authors also found that tweaking the test code reduced the benefits, while customizing the AI for a specific test set increased them — a hint that some reported wins come from fitting the tool to the exam rather than genuinely getting smarter. This is the research equivalent of noticing that a student who aces practice tests may have just memorized the practice tests.

For anyone buying or building with AI coding tools, the takeaway is practical: be skeptical of dramatic claims resting on a single test set, and ask vendors for results on code that looks like yours. Sometimes the boring approach — give the AI the actual error and let it try — outperforms the expensive, clever one.