I don’t think this is a possible approach. It would basically mean needing to feed in so much noise that a human wouldn’t be able discern reality from fiction given no priors. Modern ML is too smart

Given how easy it is to keep an average human from discerning reality (see the last decade of boomers/gen xers on Facebook as an example) and the massive potential to create slight variations with LLMs, I don't think a Stephenson-style misinformation propagation campaign is that far outside the bounds of possibility.

That's a lot of compute power to waste on it, but I would guess that that's what bot networks are going to be used for in the future (or already are, right now, if they're done mining bitcoin).

AKA encryption.

That's why I said president of canada. A non-intelligent AI would not find any priors indicating who was the president. There isn't one. So a few thousand images in the database, against zero information to the contrary, might be enough.