I’m far from an expert but as I understand it the reference point isn’t so much the “real world” as it is the training data. If the model generates a strongly weighted association that isn’t in the data, and shouldn’t exist perhaps at all. I’d prefer a word like “superstition”, it seems more relatable.