It stems from making assumptions on the human and her intentions based on the computer settings or network location. We know better than that.
Problem is design is made based on metrics and tracking, and things are tailored to the largest user base and the rest is an afterthought.
The main roadblock is the current lack of a way to ask the visitor for her language of choice in a way that does not involve words nor country flags. Tackling this in an almost universal way is not that easy and obvious, but even a partial success is better than having an automated way force a broken choice upon the human visiting the website.