I can't, as my employer owns the code, not me, but there are several examples in one of the Ruby codebases I unfortunately maintain where I can see this degeneration happen via the git history. A small 8 line method with just two parameters slowly grows in complexity over time, until one day one of the original parameters supports two different shapes, and later on it's not that easy to understand which shape it should have in the specific conditional branch you're trying to fix, and the last person to touch that code left the company 4 years ago.

The fault, of course, ultimately lies with the people who wrote and approved this nonsense, but types, or at least type hints, help to avoid this issue.

Have you been in static types the whole time? It's a really, really common failure state in dynamic programming languages, the Everything Function, that started out as something simple, but then someone added a flag to make it also do this other thing, and then you need a flag to only do that other thing sometimes, and someone needed to operate on multiple things so they made a string parameter also optionally an array, and later someone allowed it to also be an object with this one method, or maybe another method if it's present because some other team implemented that before the first one and can't switch now, and before you know it it's a free-for-all of people adding flags and options and type analysis and if statements and you have a complete mess. Especially if this function is shared by many disparate teams, each of whom isn't "allowed" to break the others, though a single team can fail this way plenty fine.

You can still do this in static languages, but they do push back a bit more because you don't get the flexibility that dynamic languages offer when it comes to accepting a huge variety of different input types.

I've torn a few of these apart over the years. Never fun. Haven't tried with AI but suspect that would only be a quantitative change rather than a qualitative change. The fundamental problem with fixing these is lack of information about the exponential complexity of possible call mechanisms and the AI will have the exact same information problems I will, just faster.

Edit: One of them that I tore apart ended up being two entirely separate functions slammed together into one by historical contingency. I don't just mean that I broke the functionality down into multiple functions, that's a basic tool of how you tear these down and is nothing of note. I mean that one of the "everything functions" I tore down had two distinct calling patterns that were distinct functions that not only shouldn't have been festooned with so many options, but never should have been one function at all because they weren't even conceptually the same thing or even particularly related.

Think of it as two stages of a straight-line process, that were just jammed together because of the fact they got called at similar times, and the original writers weren't clear on the unrelated nature of the tasks and nobody was able to see it through all the obfuscation until I sat down, very deliberately, and I realized this as I was tearing it apart. I don't remember the details, I tend to remember things very conceptually and thus I have a hard time remembering the details of functions with no conceptual purity, but you can get close by thinking of the function as validating incoming parameters, and then applying the parameters to a database. And people were so confused that despite the fact this function, when tickled correctly, could do it all in one shot, sometimes, kinda, with some caveats, there were places where this function was called first to validate (with flags to shut off the application), and then to apply (with flags to shut off the validation). And to be clear, I mean, I did not realize it either even from my contact with the function over the years. It was only when I sat down with it for hours and systematically tore it down that I figured that out.