A good classifier will take into account interactions as well, so if 80% of the people who like both Top Gear and Glee are heterosexual, then they can still get some predictive power. Further, if 90% of people from your city are heterosexual, they could take that into account as well. And that doesn't consider the internal stuff that Facebook knows (and can sell to advertisers), such as the demographic profile of the people near your IP address.
It's not just averages of society, it is individual profiling. At it's most benign, it will target ads that it thinks your demographic profile will be most responsive too, but at it's worst, it could certainly be (mis-)used for individual persecution.
The most exciting phrase to hear in science, the one that heralds new discoveries, is not "Eureka!" (I found it!) but "That's funny ..." -- Isaac Asimov