The problem with recaptcha style systems is that once the underlying model is trained, you are basically playing family feud - you aren't trying to provide a correct answer, as much as you are trying to match the "survey says..." majority opinion