Maighstir: The distorted word ones are by now so quick for a computer to solve that they're pretty much useless, they're catching up on the "identify the picture" ones too.
timppu: I quickly read the CAPTCHA Wikipedia article which mentions these machine-learning robots defeating CAPTCHAs, but in case you know more:
Is it like the robots recognize e.g. cars in any pictures (or like one MS-proposed CAPTCHA. telling cats and dogs apart), or is it merely that those robots routinely store all the images they encounter into a database, along with the right (or wrong) answers? Or do they do both?
If it is the latter, then I guess they should come up with CAPTCHA systems which don't use pre-generated pictures or stuff, but generate the pictures on the fly? Possibly there are such already?
EDIT: But then if there is a robot who can generate new pictures and decide on the fly what is the correct answer... then I guess it isn't necessarily hard for another robot to learn the correct answer as well using similar algorithms... damn this is confusing.
I only have very rough knowledge, but it's based on machine learning. Figuring out -based on earlier knowledge- what the current image most looks like (a C, an A, and a T; a feline pet; or a bulldozer), not trying to store half an infinity of images, going through them each time and being confused when a few pixels change because the checksum isn't known.
And yes, computers are becoming better than us at the stuff we specifically use to make systems computer-proof.
We teach computers to read and understand distorted and hand-written text, and then try to use distorted text to keep computers away from only-humans-allowed systems.
We teach computers to recognise myriads of traffic signs in various angles and light configurations, and then try to use pictures of use traffic signs to lock computers away from system we only want humans to have access to.
Confusing? Yes.
Google does both cases in the latter at the same time. You get multiple images, some of which they know are <thing>, and some of which the computers are not yet entirely sure is <thing>, so they ask you to confirm in a way that you don't know which is "I think this is a <thing>, but I'm not entirely sure, can you tell me?" and which is "I know what this is, do you?".