Ultimately I think humans are also saying "I dunno, but here's what everyone else seems to think is a chair". But doing so more accurately, because they're more correctly modeling what people don't think are chairs.

I suspect that there's a fundamental difference in how humans learn from positive vs negative feedback, and in particular that there's a more likely to be a symbolic step in negative feedback. When we see an exception to a rule we look at the rest of the data set to find the basis vector on which the exception is operating in a way that I suspect we don't when the data point expands but doesn't conflict with the current model.

AI could be doing this, but I don't think it is.