Can AI Fix Your Code? Understanding the Strengths and Pitfalls of ChatGPT-Based Program Repair

Introduction

Imagine you’re working on a piece of code, and suddenly—boom!—you hit an error. Debugging can be a nightmare, especially when you’re dealing with cryptic error messages and complex dependencies. That’s where automated program repair (APR) comes in. The idea is simple: let artificial intelligence analyze the bug, suggest a fix, and save you the headache.

With the rise of large language models (LLMs) like ChatGPT, APR tools have become more powerful and intuitive. Conversational repair techniques, such as ChatRepair, allow developers to interact with AI, receiving bug explanations and improved fixes through iterative conversations.

Sounds great, right? But here’s the catch—these AI-powered repair tools still fail a lot.

Researchers Aolin Chen, Haojun Wu, Qi Xin, Steven P. Reiss, and Jifeng Xuan set out to answer a key question: Why do LLM-based repair methods still struggle with so many bugs? Their study dives deep into ChatRepair’s strengths and weaknesses, uncovering fascinating insights that could shape the future of AI-assisted coding.

Let’s break down their findings.

The Promise of AI in Code Repair

Traditional automated bug-fixing tools rely on techniques like symbolic execution and pattern matching. While effective, these methods can be rigid and lack adaptability. ChatGPT-powered repair, on the other hand, offers a new paradigm: AI that can “understand” code, diagnose issues, and iteratively refine fixes through conversation.

ChatRepair, one of the leading conversational APR tools, uses ChatGPT to analyze faulty code and generate patches in one of two ways:

1. Cloze-style repair: Generates a fix for specific buggy lines or small sections of code.
2. Full-function repair: Rewrites the entire function where the bug appears, without specifying exact buggy lines.

Additionally, ChatRepair’s iterative approach aims to refine patches over multiple tries, increasing the chances of fixing bugs successfully.

Where AI Debugging Falls Short

Despite the promise, ChatRepair fails to correctly repair more than 50% of the single-function bugs tested in the study. The researchers identified three major pain points:

1. Cloze-Style Repair Often Breaks Code

• It frequently produces patches that don’t compile (about 58.9% of the time).
• Errors happen because the model introduces redundant context, makes changes outside the buggy area, or adds undefined variables.
• Full-function repair, while not perfect, produces fewer compilation errors and a higher success rate.

2. Iterative Repairs Don’t Always Help

• You’d think that having ChatGPT review and refine its fix multiple times would improve results, right? Surprisingly, that’s not the case.
• The study compared ChatRepair’s iterative approach to an alternative method where ChatGPT just generated fresh patches without iteration.
• Shockingly, the iterative approach repaired fewer bugs!
• One culprit? The AI often generates duplicate fixes instead of introducing new, creative solutions.

3. ChatGPT Struggles with Complex Fixes

• If a fix requires elements beyond the buggy method, such as data from different parts of the program, ChatGPT’s success rate drops to 45%.
• Bugs that can be solved using basic programming constructs (e.g., simple operators or standard functions) are fixed correctly 100% of the time.
• However, if solving the bug requires retrieving data from elsewhere in the codebase (outside the buggy function), ChatGPT struggles.

Why Does ChatGPT Get It Wrong?

The researchers pinpointed three fundamental reasons why ChatGPT-based repairs fail:

1. It misunderstands the root cause of the bug.
2. ChatGPT sometimes misidentifies the faulty code or makes incorrect assumptions about what’s broken.

3. When its initial understanding is wrong, it almost never produces a correct patch (0.8% success rate).

4. It doesn’t fully grasp the program’s expected behavior.

5. While failing test cases provide some hints, ChatGPT often struggles to infer what the code was meant to do.

6. This is especially problematic when a fix requires adding new functionality rather than modifying existing logic.

7. It can’t always find the right building blocks for a fix.

8. If a patch requires copying code from other parts of the program, ChatGPT often misses or misuses key elements, leading to incorrect repairs.

What This Means for AI-Powered Debugging

The findings share valuable insights for both developers using ChatGPT for debugging and researchers working to improve AI-based repair tools. Here are some key takeaways:

1. Method-level repairs work better than line-based repairs.

2. Instead of narrowly focusing on a few lines of buggy code, giving ChatGPT the entire function often results in more accurate fixes.

3. Blindly trusting AI-generated fixes is risky.

4. Even though ChatGPT-generated patches may seem plausible, they often fail to fully solve the problem or introduce new issues.

5. Developers should manually review AI fixes instead of applying them unquestioningly.

6. More context is key to improving AI-generated fixes.

7. Future research should explore ways to provide ChatGPT with more information about program behavior and structure to improve fix accuracy.

8. This could involve feeding it additional context, such as related functions and expected output descriptions.

9. Iterating isn’t always better.

10. While iterative refinement works well in natural language tasks, in code repair, repetition often leads to redundant solutions instead of improved fixes.
11. Alternative strategies for refinement should be developed to prevent duplication and encourage more diverse solutions.

Key Takeaways

• ChatGPT-powered debugging is promising but far from perfect. It still struggles with complex fixes, iterative refinement, and understanding expected behavior.
• Method-level code repair works better than pinpointing smaller code fragments. Giving AI more context helps it understand and fix bugs more effectively.
• AI fixes should not be blindly trusted. Developers should carefully review suggestions before applying them to their codebase.
• Providing more context could improve future AI-based repair tools. AI struggles with handling missing behavior, external dependencies, and complex debugging logic.
• Iterative refinement doesn’t always help. In some cases, re-asking ChatGPT for a fresh solution works better than making it modify a previous fix.

Final Thoughts

The idea of an AI that can automatically debug code feels like science fiction, but we’re closer than ever to making it a reality. Tools like ChatRepair represent an exciting step forward, but as this research highlights, we still have a long way to go before AI can reliably fix all coding issues.

For now, think of AI repair tools as helpful assistants rather than replacements for human debugging. They can provide useful insights and quick fixes, but human oversight is still absolutely necessary.

Looking forward, improving AI’s understanding of program behavior, expected outputs, and surrounding code context could significantly boost its repair capabilities. As AI models continue to evolve, we may one day reach a point where AI debugging moves from “hit or miss” to a truly indispensable tool for developers.

Would you trust an AI to fix your code? Or do you think debugging will always require a human touch? Let us know your thoughts in the comments!

Sources:

Chen, A., Wu, H., Xin, Q., Reiss, S. P., & Xuan, J. (2024). Studying and Understanding the Effectiveness and Failures of Conversational LLM-Based Repair.

Published in: arXiv preprint

What do you think? Have you used ChatGPT for debugging before? Share your experiences below! 🚀

If you are looking to improve your prompting skills and haven’t already, check out our free Advanced Prompt Engineering course.

This blog post is based on the research article “Studying and Understanding the Effectiveness and Failures of Conversational LLM-Based Repair” by Authors: Aolin Chen, Haojun Wu, Qi Xin, Steven P. Reiss, Jifeng Xuan. You can find the original article here.