The same tech works in both directions. Spammer creates 1000 email variants using a LLM, spam filter collapses those 1000 variants back into easily classifiable embeddings

Isn't this exactly the same problem as detecting if text was AI generated which has largely agreed to be impossible?

Kind of, but not exactly.

In the end it's content that matters, not the form that it's send to - I don't care if it's my grandma sending me offers for viagra pills, or a spammer, and I don't care about the language either - I just don't want to get such offers.

On the other hand, if there is an e-mail that may be interesting to me, I don't care if it was sent by a human, or by a machine - I want to see it in my inbox.

In other words - it's not about distinguishing human/machine written text. It's about distinguishing content that is worthwhile for me from the one that isn't.