The expense will down trend for cost and power of machines and thus content. Human brains however are effectively stable and have physical limits that aren’t increasing.

I can imagine even with a 10% hit rate, perhaps using public records like birth certificates or similar, there will still be economical benefit for the bots/targeted ML. Whatever a human can do in the digital world, a simulation can do better, faster, and cheaper.

I’ve read and appreciate brugeman’s approach and research. Spam typically always has a call to action - however when you can instead hyper-target, again it’s death by 1,000 cuts.. most or all don’t have a link or product they are trying to sell you. It will be selling instead a way of thinking, a reality that is one they benefit from (they being whoever breaks even or makes money).

I have a Nostr relay too and my spam ML engine hasn’t been updated in training for 5 months now. It has 99.999% accuracy or abouts. Spam is different from targeted content.

Consider a Wikipedia page with two versions and one promotes euthanasia in the death section for suicides and the other condemns and highlights doctors are mass murders. Imagine Wikipedia shows you the version it wants you to side with. It’s very different problem from spam detection.