The trick is to just use something no-one else is using. As long as it’s sufficently different and you’re not a big enough site to make you a target on your own then the spammers will go elsewhere where it’s easier.
It can be something simple as long as it’s different from what everyone else does.
The basic problem is that everything a human can do you can train/write a program to do – what you need to do is make each problem different so the rewards for solving any one aren’t worth the effort required.
Inspired by something I read elsewhere, can’t remember where and can’t find it tho’. Sorry.
When I first heard about photo-based CAPTCHAs, this was the biggest problem. Some guys at CMU were trying to build a little webgame to help better classify pictures. Basically, they’d show the same picture found on Google to several people and you got points based on how quickly you agreed on the subject of the photo.
That was a few years ago. Now there’s a better way. The photos have already been laboriously tagged and organized. Here’s the result of a Flickr search for “motorcycle”. Not all of the pictures are readily identifiable as motorcycles, but the tagging is really quite good. And the keyword search can be done dynamically by the testing computer, mixing any number of keywords. Very very hard for a bot to beat.
Of the other problems you mentioned, I think that the language issue is the most important, but it may be necessary to keep the spammers out.
Actually, that’d be easy for a bot to beat. The bot just picks out the key word, “motorcycle”, in this case, and does a Flickr search of its own, to see which images come up. Then it sees which of those images that comes up are amongst the ones to be tested.