All the falsehoods texts discuss the same kind of issue: the domain doesn't match the model. And computer models are imperfect a priori for modelling empirical data(natural language, human identity, relationships, continuous functions, physics calculations). We can get those closer in approximate so that the error is negligible, by testing, collecting feedback about our error, and iterating on the model. All of these - even the genders and natural language - can be tested for error, we just can't ensure 100%, and our measurements for the human things necessarily rest on sampling opinion.