I think the AI-is-going-to-kill-everyone hysteria is absolutely overblown, both by those who believe it, and the media covering them, but one thing that's always bothered me about the counterpoint is that it a common argument is "AI is bad at what we want it to do so how can it be dangerous?"

This imagines that the only way for AI to do serious harm to us (not even in a "kill everyone" sense) is for it to be some super-competent Skynet-level time traveling demigod. I think it's much more likely that if there is some sort of AI calamity resulting in a lot of deaths, it's because the AI just doesn't work very well and ends up breaking all the HVAC systems in a country during a heat wave or something, rather than a singularity-type event where it "decides" to actively start hunting humans.