Comment by ben_w

If the AI is so monomaniacally focused on paperclips (or anything else) to be a threat to us, going to some other planet is simply one of the early steps, but they absolutely will come back to Earth after all other resources have been consumed.

If such an AI can be reliably made to never ever come back to Earth, they were never a threat in the first place. Nobody knows how to fully test an AI's utility function yet, only randomly test inputs and hope the random distribution we chose is helpful; but every time a diffusion model's output is body horror, every time an LLM makes buggy code (and even every time it gets the pelican-on-bike wrong), this is an example of the test distribution not being good enough.