this is only one example of a large number of problems with SpamAssassin.
it is a problem with the configuration of spamassassin, not with the
software and not with the basic paradigm of rule-based filtering. it is
simply a matter of tuning the ruleset better.
true, much of the problem is with the default rulesets. they might
correleate well to observed spam but that doesn't mean they are good
predictors of spam.
a more general problem is the notion of applying rule-based filtering
on behalf of recipients who don't either understand or choose the rules.
rules can implement recipient preferences, but it's simply not true
that you can reliably determine whether something is spam by looking at
it, no matter how "well-tuned" the rules are.
Keith