Why AI Just Can't Fix that Bug

The steps between feeling like a Prompting God to feeling like you are being gas lit by an AI is roughly 7 steps.

Aug 14, 2025 - 02:00
 1
Why AI Just Can't Fix that Bug
Why AI Just Can't Fix that Bug

If you’ve used ChatGPT, or any AI in the past few years, then you probably ran into issues where it just can’t get things right. You started off at the beginning, and as you continue trying to prompt to get a fix, something breaks or you run into what feels like a lie to your face. Where it feels like the longer you go, the dumber the AI gets.

There is a reason for that. It’s called Debugging Decay.

Why AI Just Can't Fix that Bug
Model decay shown per prompting attempt to fix the problem (Adnan & Kuhn, 2025)

Most models lose 60 to 80% of their debugging capability within 2 to 3 attempts, according to Adnan & Kuhn (2025). They introduced in a 2025 paper the Debugging Decay Index, which is a framework quantifying when debugging becomes ineffective. As seen in their attempts, for example with GPT-4, where the accuracy and effectiveness is at 90.2% in the beginning, it drops to 40%, 20%, 10%, 3% and below 1% by attempt 7.

The why is pretty simple. It's called Context Pollution. Either:

  • An error was made earlier and wasn’t caught, it keeps going into what didn’t work earlier.
or
  • The AI made a bad assumption, never thinking to question it.

If the model made mistakes from the beginning, that’s within the context window. As you continue to prompt, it pulls back from the context so far, which is the conversation up to that point. Or it still made a mistake with a bad assumption, but doesn't think to question it. It'll continue behaving as if nothing is wrong, which will mean it'll continue to pop up.

????
The latter is an issue that most vibe coders will face, especially if they don't have the background knowledge of knowing if an answer is good or bad.
Why AI Just Can't Fix that Bug
Debugging effectiveness varies significantly across model architectures depending on attempts to fix the mistakes. (Adnan & Kuhn, 2025)

That said, it’s not like there are no solutions.

  1. You can try to fix the context window. “Give me the top 10 possible issues, and how to test if it’s that issue” before asking it to fix the code or design error. It puts new information into the context window that it can pull from when trying to fix the error.
  2. Include the errors done in screenshots, and restart the prompt based on the work done up to that point. Reset the context window.
  3. Give the bug to another model. Think of it as a fresh set of eyes from someone else.
  4. Do a better prompt from the beginning.

While the paper focused on debugging when it comes to coding, the same principles still apply for non-development related tasks. And the easiest solution is to just restart the context window with a new prompt.

Sources

Adnan, M., & Kuhn, C. C. N. (2025). The Debugging Decay Index: Rethinking debugging strategies for code LLMs. arXiv. https://doi.org/10.48550/arXiv.2506.18403 (arXiv:2506.18403)