It seems to me that competent use of even current-gen, non-frontier LLMs^[1] could be a totalitarian's wet dream. Things like "heavenbanning" have gone from 'this sci-fi idea might be legitimately frightening in a few years' to 'an MVP that fools most people could be built in a week by a few programmers with limited ethical capacity'. A custom model fine-tuned on a platform's data and red-teamed not to reveal that it's an AI could leave pretty much everyone doubtful of their reality^[2].

A state that can guarantee 1:1 links between real identity and website requests can neutralize the latter defense, and an intelligence community that's sufficiently interconnected with each platform can neutralize the former. Plenty of states like this exist today, and more as the world and the internet shift away from what they were in the early 2010's.

1. ^{^}

   Of the kind that would cost five to six figures to train, even if no open-source models existed - well within reach for unethical small companies, let alone nation-states.

2. ^{^}

   For now, at least, a clever dissident could challenge 'friends' to link up on another platform, and all the standard "Am I shadowbanned" techniques that any bot farm automatically runs as part of its daily work cycle will work likewise for a human.