Markov chains are the next generation of spam filtering tools
(some of the GPLed tools already do this).

It involves keeping frequencies for token sequences rather than just
individual tokens. It makes the spammers job a lot harder, though
not impossible. The downside is the increase in filters database size
as the number of tokens increases exponentially with chain length.

It all comes down to making the filters model of written emails have more
parameters and being more accurate than the spammers until in the limit
the spam email looks like a valid one (IE. it contains usefull information :-)

-Steve