Clickbait: Which types of AIs, if they work, will upset the current gameboard and make a large difference 200 million years later?
Some forms of Limited AI are so limited as to be of questionable relevance within the larger dilemma.
A [central example](central example-1) is a heavily Boxed Oracle that can only tell us whether certain theorems are provable in Zermelo-Fraenkel set theory. While this Oracle may indeed have been so heavily Boxed as to be genuinely safe, nobody knows how to save the world using a trustworthy means of knowing that certain theorems are provable in ZF.