I'm a little surprised that AI is the best way to approach this, as opposed to some kind of transpiler. I'd be more inclined to trust the latter for correctness.
But I still agree with you. I've experience of a couple of COBOL migration projects, and I wouldn't want AI anywhere near it. The testing required is insane enough as it is, and the once-a-decade corner cases numerous enough for the idea of using AI on this to scare the bejeesus out of me.
This is code that can still contain for-loops to slow down processing so the teleprinters can keep up...
As for COBOL, it was added in 2002.
"Why Move to JVM COBOL Code?" - https://www.microfocus.com/documentation/visual-cobol/vc60/D...