The article says: "Statistical (bayesian) scanning is easily defeated by randomization".

This is simply not true and I guess folks at hexview haven't met any real bayesian anti-spam application. I develope (and use!) a statistical (though not Bayesian but inverse chi square) content filter and I can tell you that it's far from "being defeated". I get lots of spam every day and it marks them correctly and catches at least 99.5% of the spam easily.

Nowdays most of the spam is sent by botnets as illustrated in the "Many-To-One" scenario. They accept that it's a difficult to handle situation and heuristic filtering is required unless the bots send a high email traffic. What about new botnets, unknown to the STP system? Bayesian filters can handle this. A typical shortcoming of this STP thing is that they cannot handle situations when you get spam from a low traffic host or if your colleauge bothers you with some stuff. Statistical filter can help you with this, too. It's unacceptable for me if a 3rd party judges my emails whether they are spam or not. That's why I keep avoiding solutions like STP, RBL, ...