ChatGPT doesn't "know" anything. It's not sentient, it's just running statistical models.

I'd be interested in seeing the prompts where it gave incorrect code and immediately apologized without user interaction. Maybe the training data includes text showing how to bad code would look like and the parts where good code appears didn't rank high enough.