So in your thesis, auditors and insurers will happily accept the intractable nightmare of mainframe code hacked together over multiple decades, often by the lowest-cost outsourced development shop and with the original creators now long retired, but they'll refuse modern, readable, testable code because it was generated partially by AI?

I've worked on the type of legacy IBM systems being discussed here. They're often filled with absurd hacks. VMs running outdated operating systems on emulated hardware because that was cheaper than updating the application. Systems so confusing and poorly documented that only a single employee knows how to update it. Known bugs in software that are solved by pre-processing or post-processing inputs/outputs to hack around the issue.

"Vibe coded slop" isn't going to be any worse than "IBM-outsourced slop" or "Maintained by one quirky IT guy for 15 years slop"

AI is good at coding if there's an oracle. If the system is ancient, unreadable, untestable, that's exactly the opposite. It won't get the exact set of corner cases.