On the Limits of Generalizability