If enough people colluded to use the same unknown word, say "foobar", then couldn't they train recaptcha to believe that is the true value of the word? If I understand recaptcha correctly, and assuming they don't detect collusion well, then eventually the known word pool would get poisoned with a surplus of foobars.
>Checking for offensive words is only one of the filters we have. Even without that filter, it’s essentially impossible to get recaptcha to return a false result.
This has happened in google image labeler. Now offline, it was a game where two random people were matched and needed to find new tags for the same image - the scoring was cooperative so points were given when you both chose the same word in the same game. /b/ of course quickly organised the jesusporn metagame... Label the first image as jesus, second as porn, third with jesus, etc. This resulted highest scores for some time.
