Theory must account for data; if elicited data are shaky, other data are supplied by introspective fiat. These data conspicuously fit the existing model. If these data are questions by proponents of another model, then the search is on for speakers, be they naive or trained, whose idiolects provide data that fit the first.
He's saying that's a bad thing, FYI. "Introspection" is basically a linguist thinking about it and pronouncing his view as the correct one.
It is in a theory-internal sense that the distinction between 'grammatical' (derived from a grammar or accounted for by a grammar) and 'acceptable' (derived from elicited or introspective evidence) is to be understood.
Birdsong quoting from Levelt et al:
The empirical domain of Chomskian linguistics is linguistic intuitions. The relation between these intuitions and man's capacity for language, however, is highly obscure.
(This does seem pretty nuts.)
Linguistics is a potentially fraudulent enterprise when elicitation data can be manipulated to substantiate pet theoretical analyses.