Imagine that a model tasked with detecting offensive text determines that whenever the word “immigrants” is mentioned, the text is offensive. Or, for example, where an image labeling model determines that only the two left-hand images are photos of a wedding, while the right-hand image is just of a couple…
