Current models are amazing but production code incorporates unknowable information through decades of diagnosis and bug fixing (and sometimes the opposite: working around bugs in other software that it talks to). Put even the most amazing coder that was ever born onto improving a complex old codebase, and you can guarantee they are going to break stuff. AI is better but not infinitely better.
People keep repeating this but on my experience is also very good at finding kirks of specific compiler versions.
I think it is a reasonable risk to encounter these bugs again in order to get a more clean and documented codebase. Especially when asking AI to plan and A/B release is also trivial.