Something similar to this is explored in Permutation City (a book), where "smarter" emails won't disclose their message unless they think they're disclosing it

Something similar to this is explored in Permutation City (a book), where "smarter" emails won't disclose their message unless they think they're disclosing it directly to you, so people train AIs to run the emails in a sandbox and pretend to be the actual human operator to detect spam, and the emails are trained to try and figure out if it's the real intended recepient before getting to the spam part, and so on.