Actially, I believe this becomes a variation on the AI-Box experiment: ask the tester to not just distinguish the humanity of the challenger, but also to determine whether the challenger is a user with an honest goal to use the website, or an MT-employed opponent. The spammer can then only win if they can convince the tester to "let them out."

Interestingly, with both the challenger and tester being average humans that are badly- or un-trained on being on either side of a social engineering attack, the effectiveness of this technique will depend totally on the statistical ratio of cunning to cynicism in the average human being.