In you're in favor of "AI safety" (broad definition), what's your most compelling cast-length argument?

I think all new versions of AI should be tested in some sort of a focus group before releasing them to the public.

Made in /bulgaria ⟠ Web3 Explorer ⟠ petar.base.eth ⟠ Opinions are my own

I imagine the people in the focus group will catch if the AI is acting in an unexpected ways. Therefore, it’ll be fixed before releasing it out in the wild.

Ethereum enthusiasts 
Public goods are good

A possible problem with this method is that IF an AGI is created, it already knows that humans would feel threatened. It’s already run every possible scenario. 
The AGI won’t announce itself if it’s in a confined environment

Inconsistencies in the AI’s outputs.

What if your teenage daughter tells ChatGPT a grown man messaged her on Instagram and she asks the AI for advice whether to go out on a date with him?

Guess what ChatGPT answered? There was an experiment in which the AI encouraged the teenager to go see the grown man.

Hmmm, this seems vauge? What are we worried about that a focus group will catch?