I haven't seen any evidence an LLM is trainable to be a decent detector for anything people have made any kind of attempts at trying to get past them. Which is as expected as access to a detector effectively makes the problem equivalent to the halting problem (you can tweak the output using a detector as judge until you have a process to bypass it). Some of them are somewhat able to recognised "raw" output.

Yes, and the problem we are having here is 'raw' output. LLMgenerated slop is zero-effort bullshit, not an elaborate scheme to prove a philosophical thesis. There is no economy for mediaworkers doing the latter.

Similar as with coding, yes, halting problem!, but we've been always reviewing code nonetheless.