Many linguists believe that a parameter-setting model of grammar should
capture typological generalizations. For example, a particular feature's
cross-linguistic rarity might be 'registered' in a grammar that possesses
that feature by means of a marked setting for the relevant parameter. I
argue that such a view is in error. Grammars do not encode typological
generalizations, either directly or indirectly. Put in a somewhat different
way, Universal grammar tells us what a possible language is, but not
what a probable language is. The most robust typological generalizations
—those arising from the seminal work of Joseph Greenberg— have an
explanation based in language processing.