Source: just left GOOG after 5 years on the GitHub tooling team.

And the fact is that the author of the web page doesn’t know the language of the content, if there’s anything user contributed. Should you have to label every comment on HN as “English”? That’s a huge burden on literally every internet user. Other written language has never specified its language. Imposing data-entry requirements on humans to satisfy a computer is never the ideal solution.

1) Pronounciation is either solved by a) automatic language detection, or b) doesn't matter. If I am reading a book, and I see text in a language I recognize, I will pronounce it correctly, just like the screen reader will. If I see text in a language I don't recognize, I won't pronounce it correctly, and neither will the screen reader. There's no benefit to my screen reader pronouncing Hungarian correctly to me, a person who doesn't speak Hungarian. On the off chance that the screen reader gets it wrong, even though I do speak Hungarian, I can certainly tell that I'm hearing english-pronounced hungarian. But there's no reason that the screen reader will get it wrong, because "Mit csináljunk, hogy boldogok legyünk?" isn't ambiguous. It's just simply Hungarian, and if I have a Hungarian screen reader installed, it's trivial to figure that out.

2) Again, if you can translate it, you already know what language it is in. If you don't know what language it is in, then you can't read it from a book, either.

3) See above. Locale is mildly useful, but the example linked in the article was strictly language, and spell checking will either a) fail, in the case of en-US/en-UK, or b) be obvious, in the case of 1) above.

The lang attribute adds nothing to the process.