Failover to Human Intelligence

"AI-written code is arguably better, often with more comments and docs, humans would understand it faster anyway"

Strongly disagree, AI is often confidently wrong and leaves bugs all over the place.

With human code you know at least when the dev says he ran it and it works that it does.

Developers will not lie about Apis they implement

We are moving up the abstraction hierarchy. Most programmers don't write low-level code in today's age. We kept moving up and now we have natural language programming. The programmers of tomorrow will mostly be using nl to program, that doesn't mean that there won't be "lower-level" programmers maintaining things.

As for the critical part, the way I see it is that AI will excel and readily be applied in low-stake environments (Frontend, Games, etc) where you can simply loop the AI agent until it clicks and gets the right output.

For high stake environments (healthcare, self-driving, etc) AI will reluctantly be used (for the foreseeable future) because the error cost is too high.

I could be completely wrong.