A number of authors have described randomized algorithms for solving the epsilon-approximate nearest neighbor problem. In this note I point out that the epsilon-approximate nearest neighbor property often fails to be a useful approximation property, since epsilon-approximate solutions fail to satisfy the necessary preconditions for using nearest neighbors for classification and related tasks.