[ademarre]

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.

[_ofdw]

Some 4channers tried that already. You can guess which 6-letter epithet for black people they tried to use.

[Ntrails]

Ladies and gentlemen, I give you: "the internet".

[hueving]

>You can guess which 6-letter epithet for black people they tried to use.

'Blacks'

[nyrina]

Hint: It starts with "N" and ends in "igger"

[wernercd]

Naggers!

[mreiland]

did it have any letters in between?

[Rylinks]

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...

[shadowfox]

> it’s essentially impossible to get recaptcha to return a false result.

Famous last words :P

[viraptor]

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.

[sesqu]

ReCaptcha does do some collusion detection, and I assume they do it well, since it has not publicly been stated to be a problem area.