it is a little funny to me that they're taking about using AI to detect AI garbage as a mechanism of preventing the sort of model/data collapse that happens when data sets start to become poisoned with AI content. because it seems reasonable to me that if you start feeding your spam-or-real classification data back into the spam-detection model, you'd wind up with exactly the same degredations of classification and your model might start calling every article that has a sentence starting with "Certainly," a machine-generated one. maybe they're careful to only use human-curated sets of real and spam content, maybe not
Ultimately, LLMs don't use words, they use tokens. Tokens aren't just words - they're nodes in a high-dimensional graph... Their location and connections in information space is data invisible to humans.
LLM responses are basically paths through the token space, they may or may not overuse certain words, but they'll have a bias towards using certain words together
So I don't think this is impossible... Humans struggle to grasp these kinds of hidden relationships (consciously at least), but neural networks are good at that kind of thing
I too think it's funny/sad how AI is being used... It's good at generation, that's why we call it generative AI. It's incredibly useful to generate all sorts of content when paired with a skilled human, it's insane to expect common sense out of something easier to gaslight than a toddler. It can handle the tedious details while a skilled human drives it and validates the output
The biggest, if rarely used, use case is education - they're an infinitely patient tutor that can explain things in many ways and give you endless examples. Everyone has different learning styles - you could so easily take an existing lesson and create more concrete or abstract versions, versions for people who need long explanations and ones for people who learn through application