Earlier quoted context omitted.
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.
An engineer said they have defenses against this >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. http://musicmachinery.com/2009/04/27/moot-wins-time-inc-lose...
Famous last words :P