A lot of this code was written before meaningful computer networks existed. Records were committed nightly to tape banks. System limitations meant that a lot of unusual situations were handled with special magic values. Like imagines the norm of returning -1 to represent an error but all over the place and stored in databases. I'm sure that some of this was avoidable, but a lot of it was natural just given that codebase is so old.

This is how you get things like Musk saying that there are a gazillion 140 year olds receiving payments and that this must be fraud when actually this is a special data case. So what happens when the codebase is rewritten to be "simple?" Those people stop receiving payments because the special case handling gets lost. All that data consistency code that exists to handle the fact that a bunch of systems don't perform atomic global updates? Lost. Oops, rows just randomly get dropped now.

We can argue about the merits of LLM-driven programming, but it should be very obvious to everybody that "rewrite the social security codebase in a month with AI" is not well suited for these tools today and requires an incredible amount of arrogance to pursue. The annoying new hire who looks at the codebase, declares that it is messy, and demands to rewrite it but x1000.