Yes, this would break it temporarily. In the long run this strategy would lose. With enough eyeballs scrutinizing a site, such a system would eventually ident

Yes, this would break it temporarily. In the long run this strategy would lose. With enough eyeballs scrutinizing a site, such a system would eventually identify a site as spam and have it marked because of the dynamic nature of human judgement. So let's say that out of 100 visitors, 20 see legitimate content, the next 20 see spam, thus alerting the system that they have encountered a spam site. The Bayesian filtering would be for sites with established spam content and then matching it against other established spam content. This would be a cost of 20 visitors exposed to spam while 20 others are exposed to legitimate content. The remaining 60 users are diverted. Sounds like a reasonable tradeoff to me... Spammers could keep on shifting the content of sites, and this would make it more difficult to apply the spam filter for similar content (but humans, the first line of defense, aren't so easily fooled); this is where site moderators can identify users as potential spammers and give them the boot. So instead of having a 1:1 filter for domain:content, you could have a 1:n filter and then match that against other sites with similar 1:n "content shifting". A danger here is that legitimate content gets associated with spam and thus gets filtered out; this would require more sophistication on behalf of the Bayesian filter in associating content - but I'm certain this is a solved problem that can be put to use here.