There are ways to figure out language with very short text. In fact,
one can identify language changes in documents that contain text in
multiple languages.
http://www.stanford.edu/class/ee380/Abstracts/090114.html
That's not to say that Twitter uses such methods, just that it's
possible to
The language selection is useless, even with a limitation to English.
The problem is probably that normal methods of attributing language
are more or less based on longer text - and not text stripped down
to 140 chars or less.
If you want to make detection f.e. in search, rather get all
results
