Read this before and basically, they've got past 90% of Captcha tests (You know the "Prove you aren't a robot" Text boxes on online sign ins) It's AI seems to be able to learn and think like a human. http://www.bbc.co.uk/news/technology-24710209
I always thought intelligent written or spoken questions that required common-sense answers would be the best human test. Then SIRI came along :|
When I signed up for the FreeNAS forum trying to get an answer to an issue, one of the questions was "What does NAS stand for". Google can answer that one, some of the users... I'm not so sure about. Also: http://xkcd.com/810/
I don't think captcha necessarily have to be unbreakable, they just have to be more trouble than it is worth to crack. They may be eminently crackable from a technical standpoint, but whether this is practical from cost/benefit pov for nefarious types is another question. Advanced AI is expensive and captchas are relatively easy to refine. It's like putting a decent lock on a bike. It wont stop a determined thief, but it may make your bike a less attractive target than others.