Because you don't (and probably can't) know that this model is accurate? Also, there are a lot of populations which are not present on Twitter at all, and more which are not present in meaningful numbers.
Such models presumably can be tested in some sense. E.g. build some sort of predictive model of elections in each state using Twitter users as a surrogate, see how well you do.
I'm not claiming that such a model is guaranteed to be perfect or even as good as a pollster, but it sounds in principle like a reasonable thing to do, if you take appropriate precautions to reduce bias.
I'm not claiming that such a model is guaranteed to be perfect or even as good as a pollster, but it sounds in principle like a reasonable thing to do, if you take appropriate precautions to reduce bias.