My guess is that it'll be like "Nigerian prince" scams: it's not all that convincing. But even if it only works on 0.1% of its targets, that's still a large absolute number of people.
That would also be the case for just spamming people with "go join our cult" links, and I haven't seen this once. The limits load on other things:
- Local presence of the cult
- Ability to operationalize victims (convince them to do terrorism, in your model)
- Ability of victims to "function" in society (hold down a job, do cult administrative stuff, earn money)
For an idea of what is the upper bound, look at Scientology or Jehovah's Witnesses. They don't have to use an AI to talk to people, they just hand out fliers on the street and talk to whoever is interested. Lonely people already today can sign up for Scientology, so the AI adds nothing.
- Local presence of the cult
- Ability to operationalize victims (convince them to do terrorism, in your model)
- Ability of victims to "function" in society (hold down a job, do cult administrative stuff, earn money)
For an idea of what is the upper bound, look at Scientology or Jehovah's Witnesses. They don't have to use an AI to talk to people, they just hand out fliers on the street and talk to whoever is interested. Lonely people already today can sign up for Scientology, so the AI adds nothing.