That's likely not a trained behavior, though, it's probably the result of filtering the training data. It's not "when these parameters fire, trigger a refusal", it's the absence of parameters triggering the flinched words in the first place.