When knowing what's easy is hard

A funny thing about software for smart , non technical people, is that it's not always easy to know what's incredibly easy to do with code and what's astonishingly hard, or impossible.

About 12 years ago I was part of making a "second screen" app for a big TV show, it was hugely enlightening.. It turned out that moderating comments on Forum was almost impossible, but processing payments or adding real time chat was super easy. 

I never became close to being an expert, but you started to get a grip on some of the principles, like API call outs, or abstraction layers, or composability, rather like a laymen could look at a building and get a vague sense of that something looks hard to make or would defy gravity or material science. 

The hard thing about AI is even after learning the principles it seems super hard to know

- What is easy

- What is hard

- What is error prone

or even what is happening.

If I ask AI to generate a joke ( from easy to hard) 

-is it just finding an existing joke? 

-is it just rehashing existing jokes?

-is it understanding the logic and rules of humor and making a joke,?

-Is it conceptually aware of humor, and authoring a joke? 

-Is it a sentient like thing that is able to create jokes and laugh at them (no)

I continue to be stunned at how good AI can be at things which seem almost impossible, and staggered at how bad it is that seem to be easy, or hard to get wrong.

It would be super fun to gather a list of the most and least magical things and try to understand why.

How is it able to tell which small Italian town I'm in, merely from a picture of a roof and vegetation, but how is it incapable of editing a picture to add stairs to a house. 

The list goes on and on.