Unless you're pulling humans out of the streets at random and asking them questions or to do work, I guess you also shouldn't do that with statistical models of random human language.