Here’s another weakness of empirical inductive statistical observation: it can’t define the range of data to exclude obvious exceptions, exceptions “that prove the rule” and degraded data. Consider a stuck key on the keyboard. The repetition of letters on the screen has to be included among the behaviors of the key strokes. So the statistical account of the relation between keys and screen letters has to indicate that each stroke can result in multiple letters on the screen. At best it indicates an aberration, but the anomaly doesn’t rise to an indication of failure. It’s just a falsification of the inductive inference. Uninformative.
If you have a description of the machine independent of the stroke-letter correlation, you can systematically rule stuck key behavior out of the data set of machinery-ideal behaviors. The exceptional repetitions are now “mistakes” derived from machinery malfunctioning.
The same applies to linguistics. If the linguist, following Bloomfield or Diver, takes English to be all the utterances of English, how rule out stuttering utterances, stumbling utterances, half-finished sentences? What about utterances spoken by non native speakers who barely know English? How can the empirical purist rule out those speakers?
The generativist has a principled answer. Her account of English is based on a description of an independent phenomenon, the brains of English speakers. The account of those brains in turn depends on an independent description of brain evolution. And it goes like this:
1. All normally developing humans and only humans have this kind of language. Therefore, it is a species trait.
2. Humans learn language without being taught it — they learn it just by exposure. Therefore it is an instinct (via imprinting, “imprinting” broadly defined, along with internal extrapolation and selection).
3. Humans learn language during the maturational period. Therefore native language can be roughly defined by imprinting, selection and extrapolation during the maturational period.
From these three, it follows that only native speakers of English have command of the natural language. And since comprehension is part of the language capacity, the degraded data can be described as mistakes, the utterances of non native speakers can be excluded, and there is no circularity in defining the speakers (the circularity of the definition of English speakers was an embarrassment of the empiricist).
Hempel and Goodman point out the importance of counterfactuals and subjunctive conditionals for laws. But neither of these establish the authority of a law. Of course it’s true that a law provides counterexamples and subjunctive conditionals that mere statistical descriptions can’t. But what justifies the application of the counterfactuals and conditionals? No mere inductive generalization can support a counterfactual beyond mere conjectural hypothesis. What licenses the law is the independent background theory — the description of some other phenomenon that defines and determines the explananda. Newton’s laws of motion are a theory about bodies, not about special planetary motions. The theory of bodies explains the planets and at once categorizes planets as independent bodies like all other independent bodies, not special objects set in Aristotle’s rotating spheres. The independent background theory explains and redefines the phenomena, categorizing them and identifying those that belong and why those that don’t don’t.
If the key were to get stuck, the letters would repeat on the screen. If an utterance were spoken with prepositional phrases embedded with cross relations over its embedding, it would be incomprehensible. The machine explains.