For what it's worth, as the person at the top of this thread, I actually do take AI risk pretty seriously. Not in a singulatarian sense, but in the sense that I would be quite surprised if AI weren't capable of this stuff in ten years.
Even the oracle version is already really dangerous in the wrong hands. A totalitarian government doesn't need to have someone listening to a few specific dissidents if they can have an AI transcriber write down every word of every phone conversation in the country, for example. And while it's certainly not error-proof, asking an LLM to do something like "go through this list of conversations and flag anything that sounds like anti-government sentiment" is going to get plenty of hits, too.
Even the oracle version is already really dangerous in the wrong hands. A totalitarian government doesn't need to have someone listening to a few specific dissidents if they can have an AI transcriber write down every word of every phone conversation in the country, for example. And while it's certainly not error-proof, asking an LLM to do something like "go through this list of conversations and flag anything that sounds like anti-government sentiment" is going to get plenty of hits, too.