Paul Graham, the author of an essay that popularized Bayesian spam filtering, is keeping a list of tools that implement the technique. Jeremy Bowers has written today about why Bayesian filtering will make things worse, but I think he's too pessimistic. A good filter will be personalized based on each user's own spam deletions, making it much harder to beat, and if spammers adopt generic language just to get past the filters, their response rates will drop considerably.
*chuckle* I actually had a paragraph in there about the possibility that if the spammers were stuck with a lower-key sell, they might lose the response rates, and stop sending the emails... but it occurred to me to wonder about the sub-.1% returns, and I doubted that those people were going to stop responding no matter how-low key it was. Still a good outcome, though.
Even before Bayesian filters become widely used, spam response rates are already miniscule. I can't recall the source, but a recent news story about a spammer quoted only one profitable deal: The sale of spam services to others. Every other deal to advertise a product was barely enough to justify even a spammer's time. If they have to water down the sales language just to avoid a bunch of different Bayesian filters, it could be enough to knock a bunch of spammers out of the business.
An equally valid way of looking at it, statistically, is that the responders are the ones who are going to respond to any such offer. I don't mean to be cruel but when we're talking about fractional percentage responses, we're probably talking the mentally enfeebled or others who do not have good judgement. (Esp. for those who respond more then once.)
For instance, my great-aunt, who in her youth was a perfectly normal person, became senile but not-quite legally incompetent in her old age. As a result, she was preyed upon by many, many different organizations for "donations" that no normal person would have given. (She was literally giving to both major political parties because they both promised that the other party would kill Social Security if she didn't.)
If those are the kind of people who are responding to spam, the only way to cut response below the already rock-bottom level is to prevent it from getting into their hands in the first place. No matter how low-key it is, they may still respond.
I'm not pitching this as gospel truth; it's a theory I don't even know I "believe". But it's certainly possible, as is the theory that this could make it unprofitable enough to stop.
But note the "unprofitable enough to stop" theory is also based on second-order effects, which IMHO makes it a more reasonable theory then the filters working all by themselves.