I don't know if it is the case nowadays, but I seem to recall that in the old days they would check and if their model was highly predicting a language different than the one you were claiming they would serve the language they predicted.

Which, even if they don't still do this, I think makes total sense and is what I would do because obviously programmers make mistakes and if your page claims to be English but your language has a high chance of being Armenian and a low chance of being English I would consider it was Armenian.