And what's the certainty of this fix no introducing another bug?

Then you realize tests are failing but now you are not sure if they actually test against implementation or if they were genuinely good tests.

It's a slippery slope that adds more and more cruft over time and LLMs are susceptible to missing important details.

No certainty just like we don’t have certainty when we do it ourselves. What I’m saying is you can audit and use the LLM to help you audit efficiently: finding code, explaining logic, visualizing things etc. it’s just as powerful a tool for understanding as it is for generation