They've been using this for Google Books, which are images presented in PDF format, for quite some time. Like any OCR, accuracy depends a lot on the font and the text. If the text contains standard dictionary words in a relatively clean font (not old paper with strong serif fonts) the results are decent. Otherwise, poor.
I'm curious that this is news, because PDF scans from Google books have been text-searchable for a long time and you have been able to view the text version for at least a year. It is these text versions that have formed the basis of the book SERPs and, with universal search putting book results in the general results, OCR of text image scans have been showing up in the general search results for quite a while too.
I guess this announcement means that they're expanding that usage beyond Google Books to documents discovered "in the wild"? Or is this something different?