Pattern recognition is our best way to express what neural nets do, though no words fully suffice. It's a phrase that goes back to the time when AI was based more on propositional logic, and was used to point to what that approach couldn't do. You could use "sorting by complex characteristics" if you prefer. It's a way of pointing to complex calculations that chug a lot of data, and don't analyse (literally "break apart") well.

Paypal hit the similar barriers for a fraud-sorting task, and even in that day they found that a hybrid approach, part AI part human solved the problem well enough for that company to surive.

It is entirely possible that a couple decades from now we'll have a range of words to cover this whole territory, but I don't think we do now (or you'd have used 'em and cited 'em, or I'd have encountered them.) That we don't yet is understandable, it's early days. Right now if we have a better word or phrase (goal is kinda self-referential) I don't know it - even if I'm talking about the pattern of balance, compostion, color choice and line that makes "art," the pattern that can be called "artistic" is pretty much the best I can do. Feel free to put forward a better word if you would. DALLE-E and Stable Diffusion do seem to be able to recognize and spit out artistic results (rather than ugly, inartistic results) to a surprisingly large extent.

There are a billion ways AI moderation could be done. Many will be worse than useless as you say. One thing I notice a lot is that bad recognition systems merely sort out whatever is odd; if you are eccentric (perhaps because you are ridiculously well-educated) or highly creative you may get sorted out and punished by crap "algorithms." They're cheap-like-borscht systems, they don't give a damn about recognizing or addressing edge cases. It wouldn't be too hard to I sense you've run into a lot of that too. That's one reason why you want both far better (more computationally expensive) nets/associated logic and at least one human in the loop. But as well, as stated, spending more and getting a more complex net that is actively looking for patterns of language that suggest mere (safe) eccentricity or over-education is a case in which more and more detailed pattern-recognition is called for, not less recognition of patterns.

It's absolutely true that in the past companies (looking at you Google, Facebook) have been primarily motivated to reduce the expense of moderation no matter how crude the results. Just to take expensive humans out of the loop, no matter what breaks. That gives exactly the crap results you cite.

But I think Elon is betting that much better AI hardware (from Tesla) and more sophisticed, expensive nets (fed much more data) can, together with humans, do a far better job at a big but affordable price. Part of that bet is, I think, that computing power will decline in price steeply over the next decade and more, so it's worth spending far more (in the short run) on the moderation task, to get it right, than has been spent by social media companies to date.

He wants a lot fewer posts killed, but the bot and fake nation-state-actor posts killed, etc. Hard task. I doubt he thinks that's a dead-easy task.