The thing humans have, that so far trained machine-learning models seem to lack, is a fairly reliable BS detector. I think the best example of this is an inkblot test. Sure I can look at an inkblot and see fanciful shapes, maybe a chair, maybe two rabbits kissing, whatever. But in the back of my mind, I also know perfectly well that I'm looking at an inkblot; I can entertain all the other classifications, but would never accept them as fact.

The question becomes, how do we give an AI the concept of a BS detector? How can we train it to know, somehow, that the classification it has just made with a high degree of confidence makes no sense? This seems like a very difficult problem to solve.

Yes. Flashing lights at the correct frequency and intensity can cause a noticeable fraction of the population to immediately seize. Emergency vehicles are a known trigger.

Less glibly, yes. There are particular arrangements of roads which are known to get people killed by cars. For example, an intersection at a particular angle with particular speed limits cause cars to stay hidden in each others' a-pillars as they approach the intersection. There are bits of highway that are known to cause human drivers to drift; this is why many stretches of highway now have textured strips down the side to notify drivers that they've moved out of their lane.

Humans can hallucinate and are susceptible to optical and auditory illusions.