If AI wrote the translated code, maybe it can be used to maintain it. My experience with code translated by tools from one programming language to another is that much is lost and the result is very difficult to understand. In particular the "gestalt" of the actual written code, the semantics of the language used by the code, and the associated comments gets lost in machine translation.
But maybe this would work, because AI can either establish/maintain, or sufficiently reproduce, sufficient understanding. But of course, HOW WILL WE KNOW? Or do we just have to trust that the AI got it right?
In my experience teaching Ada to COBOL programmers (a long time ago), strong typing was easy for them to understand, they were used to reasoning about values and even the operations on those values. 'scope' was really difficult, pretty much everything is 'single global scope' in COBOL. Object Oriented was a bit of a stretch, some got it, others didn't. Concurrency was as difficult for them as it is for pretty much anyone who learned on a sequential programming language (one without concurrency primitives.) So it'll be interesting to think about how AI can 'grok' the code to add typing, scoping/object based modularity, and concurrency. (The latter is important since so much COBOL now is in transaction systems where performance is really important, and concurrency is a key part of achieving performance - the same way that modern DBMSs support concurrent access through locking, etc.)