I did a lot of tests with it (in my dev environment only), and I've also noticed that a lot of spammy *content* was not intercepted.
So I also had some doubts about it
After a research on internet I found that their algorithm is "secret", but it looks like that it does not consider only the "content" of a post, but also other factors such as the IP address from where the post comes. For instance, if your IP is not on its black-list AND it's the IP that is invoking the Akismet's API, the algorithm could conclude that the poster is trustworthy (so the return value is False).
However I think it has other main issues too (that I've already submitted to the author )
I think however I'll give it a try