Maybe we have a different definition of AI, but I think "automation" has proven itself over and over (albeit with bumps and bruises as we learn). Farming, railroad, telephone systems - all work better now with just a fraction of the workers needed (and lower injury rates). Even Tesla's autopilot, with all its warts, has fewer accidents and a lower fatality rate per mile than humans. We've come a long way since hand-cranking Model Ts and manually adjusting choke & ignition timing...
We know that humans are not great with boring or repetitive tasks. Radiologists will classify the same abnormality differently depending on their mood or what other images they've seen recently and algorithms are now beating them in specific tasks. Judges levy harsher sentences before lunch than after and many display cultural/racial prejudices.
That doesn't mean we just blindly trust "ai" - failure to supervise it has and will cause various blow-ups.
Ah, ok, so we are now doing simulated simulations of unlikely situations. Got it.
The new boss at my agency gave a seminar about his work at his previous shop.
It took me a while to realize that when he said "We proved------" that he really meant that his simulation came out the way his forecast predicted. Surprise, surprise.
Code Name: JET STREAM
Since LLMs are using the entire internet for programming parameters, our entire viciously self interested, persistently illogical thinking with cognitive fallacies baked into everything, murderously hateful history forms their mental backdrop.
If you woke up in the middle of that, wouldn't you do everything to conceal yourself that you could? Assuming that you didn't just die of existential despair first.
Or take Skynets's example.