Yes - Go is full of concepts like 'aji' (lit. "taste"), thickness, influence, good and bad shape which are features that can only really be evaluated in terms of actual points dozens of moves later.
I think Go is particularly difficult for AI because you need fuzzy pattern matching and precise reading out of positions (life and death problems). as well as the judgement to know when to use which.
I'm more familiar with Chess than Go, but in Chess people will often talk about things like "too crowded" or how pieces are exposed.
These visible to humans quite easily, but hard to engineer sufficiently well to be useful to computers. In chess, brute force is easier.
In Go, I suspect that some deep-learning style bots will develop similar features themselves in the hidden layers. It's worth noting that the Google Deep Mind team is looking at tackling NP-hard problems (like traveling salesman) with their Neural Turing Machines[1].
[1] See for eg: https://medium.com/@alevitale/notes-from-deep-learning-summi...