Jeremy Bowers writes on the hidden dangers (original) of Bayesian spam filters. Core of the argument: spammers can use any possible filter mechanism to fine tune their spam, and since the Bayesian filter is the best we have, once it fails we’re doomed.
However, if you’re trying to sell me something, you have to either a) market it in the body of the message, or b) give me a URL to look at. Here’s the simple algorithm for filtering spam with URLs in it: if the sender is in my address book, let it through. Otherwise, mark it as possible spam. Jeremy neglects to consider the possibility of personalized filters which by their nature can’t be duplicated by spammers, since they rely on information that only I have.
...