[oh-kumudo]

For Captcha, you are pointing out the one or several pictures that doesn't belong to certain category, so that seems to me the service needs to know the answer beforehand, doesn't it? It could be used for QA purpose though.

[agnokapathetic]

Don't need ground truth for crowd-sourced annotation just need agreement between n-different users.

1 randomly present captcha to few known-safe/low risk users 2. Compare answers with answers from higher risk user's Captcha responses.

[iforgotpassword]

Back when the captchas just showed two words, one of them was already verified while the other was not. So if you got the former wrong you failed, but for the other one you could type anything you wanted. The modern ones most likely work in a similar way.