In theory we don't need testers for program testing, we just need an AI to do it. But it is easier to say it, than to implement it.

I suspect it is easier to automate the programmer than the person doing/defining the tests.

I would argue that "programming" in essence is first and foremost the task of defining a set of actions or functions you want a computer to perform. This can be highly detailed (like assembly) or quite abstract (like SQL), and with AI you may even end up being able to just use English.

You will still need to check that the output matches your expectations, though.

I was thinking about automation of programming and I have usually come to a conclusion that when AI will crap out X thousands lines of code, how are you going to debug it when it does not do what do you want it do? How are you going to debug code which you did not write so you don't exactly know what is going on?

Additionally when you point out to AI what is wrong, how can you be sure, that AI has fixed the root cause or just made some fixes about that specific appearance of the bug, but it means that bug will pop up somewhere else?

I agree, and I didn't mean to imply that it's easy to automate programming, just that it's even harder to automate the creation of test criteria.

Then again, these things don't need to be all-or-nothing. If you consider GitHub Copilot, and where it can be 10 years from now, I would not be surprised if it can double the productivity of most programmers by something closer to pair programming than a code review process.

The anecdote presented by OP is basically a sequence of questions followed by responses. It's only after running the sequence and still having to resolution you'd need any form of AI or a human.

Making a case for middle managers isn't done by providing a sequence which can be fully run by a chatbot in practice.