I (with Eli's help) posted the URL http://www.internz.com/SpamBeGone
I did not render an opinion at the time, but now I can.
It "appears" (I haven't looked at source code) to base its classifying
solely on the subject line and email addresses. Since spam often has
forged envelope information, and since a message discussing a spam could
have a similar subject line, I would not be surprised at
mis-classifications.
Not using certain characteristics in other headers that are 100% accurate
in identifying spam, appears to be another deficiency.
Finally, I trained it on about a hundred messages, with followup training
necessary on about half of the subsequent messages. This false
classification rate has in general neither improved nor degraded except on
one mailing list. After the first two days of training, SpamBeGone
correctly classified most of the messages in a particular mailing list.
But strangely, for the last three days, in spite of continual retraining,
it has insisted that most of the messages in that list are spam.
As always, "your mileage may vary." It may be that I am judging messages
by factors SpamBeGone does not look at, thus confusing it on the factors
it does look at. Its Web page "Stats" certainly has an impressive success
story.
There is another tool from New Zealand, WEKA (Waikato Environment for
Knowledge Acquisition) that offers a lot more flexibility in artificial
intelligence detection approaches (including as one of several options,
the algorithm that is in SpamBeGone, but the setup and learning curve are
so much greater, that I don't feel like even trying it.