The accurate inferences that can be drawn from subtle linguistic attributes should freak you out more than they do.

Switching one good synonym can send the model off an entirely different direction in response, or so I’ve observed.