I have been noticing that Apple Mail has been marking valid mail as spam pretty often, and it doesn't seem to be improving when I train with these mistaken spam messages.
I had been thinking that Apple's Mail was using bayesian analysis to determine the spam messages. But I just found out that Mail uses Latent Semantics analysis for spam filtering. This post has a great description and comparison of Latent Semantics and Bayesian Net.
[via Stephen Hebditch's blog]