The Tom Sawyer kind, that is, known formally as Google Image Labeler:
You’ll be randomly paired with a partner who’s online and using the feature. Over a 90-second period, you and your partner will be shown the same set of images and asked to provide as many labels as possible to describe each image you see. When your label matches your partner’s label, you’ll earn some points and move on to the next image until time runs out. After time expires, you can explore the images you’ve seen and the websites where those images were found. And we’ll show you the points you’ve earned throughout the session.
The agreement requirement is clever. Of course, ol’ Tom was able to watch his friends do his work for him. Google has an even better idea — make the friends watch one another. Since they’re anonymous, they can’t collude in advance. Since they have only 90 seconds, they can’t devise a signaling system during the work period.
Update: Via Search Engine Watch, I learned that Google is using a technique it licensed from CMU faculty member Luis Van Ahn. A paper describing this “ESP game” also exists [PDF], as does a lecture Van Ahn gave at Google.