If it asks me to identify traffic lights, for example, I also include a couple of squares that aren't traffic lights, like a tree on the side of the road. I'm more selective than a random algorithm, so that correctly gets me labeled as a meatbag, but I'm definitely not providing good answers, either.

At this point, to be honest, I'm really not sure about the effectiveness of this mechanism anymore. It seems to me like a poorly-trained neural net with convincing human-like interaction (e.g. clicks spaced unequally) could produce exactly slightly-incorrect results that let me through. There's no way a smart 17 year-old wizzkid hasn't figured it out already and getting rich selling it to spammers in onion space.