a couple of years back I worked for a company that packaged AI with their sales software - it'd scan emails and roughly flag them for salespeople. I curated sets of data to feed the algorithm - f'rex, if an email says "go fuck yourself" you and I know that's a "do not try to sell me your product," but without that input the AI's got no context and guesses.

They were ultra super ethical about this, in contrast to how AI's come to mean "we can freely raid your images and text."

And I came to the conclusion that AI would cause WWIII not because of actual autonomous malevolence, but because it'd come to correlate an increased level of hostility with say, specific sandwich orders, and one evening it'd pick up on a lot of those sandwiches being orders and decide launching missiles was the next logical step.