Another link I could have sworn I posted here months ago:

http://www.zompist.com/chance.htm

Zompist editor Mark Rosenfelder lays out a simple statistical model for answering the question, “How likely is it that words in unrelated languages will resemble each other in sound and meaning?”  Of course, to answer this question one must first ask what counts as a resemblance.  Rosenfelder gives some rather amusing examples of “Proto-World” theorists, Greenbergians, and other linguistic cranks who seem to consult extremely lax criteria in deciding whether they will declare words to resemble each other.  He calculates that, under the criteria he finds in the practice of Joseph Greenberg and others, there is an 80% likelihood that any two words chosen at random from the languages of the world will be found to resemble each other.  Stricter versions of those criteria still yield uselessly high rates of similarity.  Rosenfelder closes with a killer argument supporting the claim that what we most need to determine the historical relationships among languages is not a list of word-pairs, but a table of sound correspondences.  In historical linguistics, it is the failure to produce such tables that separates the cranks from the scientists.