That's an interesting thought that every time we introduce novel classification schemes to resolve humans from computers we feed a large training set for them.
I never understood how that works, though. If I get a captcha street address wrong then that means they already had the answer, so how am I contributing?
For the street address ones you generally have to answer two questions, one that they know and one that they don't know. You're only 'tested' on one of them, but you don't know which. Once enough people have gotten the known one right and given the same answer to the unknown image they move the unknown image to the known pile.
They already know the correct answer if they are only showing you one image to solve.
When there are two images to solve, they know the correct answer to one. The other is shown to thousands of people, and eventually it is solved with high confidence.
I never understood how that works, though. If I get a captcha street address wrong then that means they already had the answer, so how am I contributing?