Translations are one of the poster use cases of Machine Learning, yet they rarely work. While everyone (sans publishers) were aware that "books" include too much context and nuance for a statistical model to handle, even short single-sentence Internet titles are too much. I'd think that they would get this right. Internet after all is a global place, where you have terabytes to extract from about every imaginable meme or short lived joke. There is very small number of local things which are traversing between language, so if it's English, it's most likely the same base of context everywhere. As we now see, there is too much local context to something like subreddit or Google Youtube niche for it to be of any sort of usefulness. I'd say that they are worse than useless, as they can give false information. Sadly, all that matter is a click. Even misinformed.