[4sak3n]

I couldn't agree more.

The obvious quirks which are easy to pick up and identify, like em dashes or overuse of a specific rhetorical device, are also easy to adapt and change. The things which consistently helps me clock LLMs are the deeper lack of cohesion, thoughtfullness, ideosyncratic quirks and just plain interesting writing and these lacks are, in my opinion, fundamental to the technology and no amount of post-training or RLHF will be able to replicate that.

It's the long tail of genuine human communication, the 0,0001% of training data which is so niche or idiosyncratic, which gets discarded in order to compress and encode the rest, which is what people subconsciously pick up on. Most of the hallmarks which people are so quick to point out, the em-dashes et al, are adressable with tweaks to the system but that lack cannot be because it is a fundamental weakness of the technology.