I’m way behind on this, but James Seng has also released a comment spam filter.
Skimming the comments it looks like a rich discussion, which I hope to read more carefully after putting the baby to bed.
James has also incorporated one of those image recognition thingees. I hope he explains how this was done somewhere. Update He explains it here.
Also, if nobody’s doing this yet, maybe the right way to apply these quasi-AI approaches is to put the positives into a “needs approval” queue while letting the rest go live. That way false positives aren’t lost and real positives aren’t live.