Earlier quoted context omitted.
I think the burrito analogy is flawed... searching for a word Google doesn't know about is more akin to ordering something that wasn't on the burrito menu at all, or say... if the cashier misheard your order and thought it could have been either chicken or beef so they made an educated guess and served you a Beef Burrito at the same time asking if that was what you meant, giving you an option to receive your correct…
But Google did show the correct results when I clicked on the link suggesting I was searching for the actual string. So it was on their menu all the time. I also communicated my request correctly to the other end; the other end deliberately changed my order!
the menu contains almost every word in almost every language that's online (plus many misspellings and made up words and lists of meaningless character permutations and...), so if you are going to automatically correct a query, some set of words that are on the menu are going to have to be excluded.
On the other hand, if we accept that this case was suboptimal, the solution seems to be one of:
1) never autocorrect, only offer a "did you mean." I would bet this would result in more people having to click than the current situation (in spite of the annoyances reported in this thread).
2) be more conservative with corrections and offer a "did you mean" more often. 500,000 results for "uncollege" and the fact that correcting it to "college" repeats a word in the query string is a good argument for this particular case, at least. On the other hand, testing for "[some random word] and uncollege" never seems to correct for uncollege unless the other word is education related, which is interesting. While the autocorrect isn't smart enough to notice that it's correcting to a repeated word, it is doing some kind of clustering and deciding you probably meant college in this case.