Today, it reminds me a study from Cmabrigde (http://www.mrc-cbu.cam.ac.uk/~mattd/Cmabrigde/). The idea is that a human needs only few letters in order in a word to understand that word (this is not okay for every word, but it should not be hard to find them).
So the idea is basically to create a captcha as an image with a word, but the word would be disordered in a way that human can read it such as:

CNOTNENT
MANAEGR
KITHCEN
etc.

Okay, based on a current OCR based attack bot, it's doable if you have a dictionary then use something like the levenstein distance and try to minimize the distance with the current word in the dictionary and the word you found with your OCR.
But well, the captcha has not necessary one word...
The only problem I can see with this method is that the dictionary you use to generate the captcha should be in the language of the targeted human. But well, for most of the websites, you know what readers/users you have...

If I have time I'd try to create a lib for this...