I started thinking along those lines back when the question online was "Is the dress white/gold or black/blue?". I still feel that there were actually 2 pictures sent out, to see how people responded to not only the question itself but how people would react to someone else saying that they were wrong, that the evidence in front of them was "wrong" (because they were seeing two different pictures).

Then in 2018, there was the Hawaii "nuke alert". How would you teach an AI about response to natural or man-made disasters? You can't have a lot of people in on it (like one of those training events for EMT/fire response) because you wouldn't get the "true" human reaction. Too many people knowing what was really happening would skew the results.

You could develop quite the database of human response to all kinds of things.