Consider that we have fairly decent anti-spam measures which do not look at the body of a message. To these methods, it is irrelevant how cleverly crafted the text is.

I reject something like 80% of all spam by the simple fact the hosts which try to deliver it do not have reverse DNS. Works like magic.

E-mail is reputation based. Once your IP address is identified by a reputation service as being a source of spam, subscribers of the service just block your address. (Or more: your entire IP block, if you're a persistent source of spam, and the ISP doesn't cooperate in shutting you down.)

To defeat reputation based services driven by reporting, your spams have to be so clever that they fool almost everyone, so that nobody reports you. That seems impractical.

How AI spammers could advance in the war might be to create large numbers of plausible accounts on a mass e-mail provider like g-mail. It's impractical to block g-mail. If the accounts behave like unique individuals that each target small numbers of users with individually crafted content (i.e. none of these fake identities is a high volume source), that seems like a challenge to detect.