James Seng, developer of the spam blocking “captcha” plugin for movabletype, also has a Bayesian filter plugin that tags potential spam comments.
This seems to work a little aggressively: I would think a post where the word “spam” appears would have a low likelihood of being spam . . . .
I guess it’s still in training mode if the comments are being published, even if they are labelled as spam.
[Posted with ecto]