[...] doing what it is programed to do. New top level goals won't just emerge, they would have to be programmed.

It might be programmed to produce new top-level goals.

("But then those aren't really top-level goals." OK, but then in exactly the same way you have to say that the things we think of as our top-level goals aren't really top-level goals: they don't appear by magic, there are physical processes that produce them, and those processes play the same role as whatever programming may make our hypothetical AI generate new goals. Personally, I think that would be a silly way to talk: the implementation details of our brains aren't higher-level than our goals, and neither are the implementation details of an AI.)

a facility administrator program [...] it will never decide to become a poet instead

For a facility administrator program to do its job as well as a human being would, it may need the same degree of mental flexibility that a human has, and that may in fact be enough that there's a small chance it will become a poet.

It will only ever do what it was programmed to.

And your brain will only ever do what the laws of physics tell it to. That doesn't stop it writing poetry, falling in love, chucking everything in to go and live on a commune for two years, inventing new theories of fundamental physics, etc., etc., etc. (Some of those may be things your particular brain would never do, but they are all things human brains do from time to time.)

And, for all we know, a suitably programmed AI could do all those things too, despite being "only a machine" deep down just like your brain and mine.