I’m not talking about the em-dash, which is not a great indicator IMO but the horrible overuse of binary oppositions with a kind of false surprise, e.g.:

The problem was not em-dashes — but binary opposition!

That sort of thing.

It is a much clearer marker of llm use than the em-dash. The sad thing is when searching for info on this the most convincing reply in search was generated by an LLM, which went on at length about why LLMs do this as some sort of consequence of their internal structure. I have absolutely no idea if that’s true — it really sounds a bit trite and exactly the kind of thing LLMs would confidently assert with no basis. I would want to hear from someone working in LLMs, but their blogs are probably all generated by an LLM nowadays. So this conundrum is a good example of a question where LLMs actively work against clear resolution.

This is in my view the most insidious damage word generators are inflicting on our culture — we can no longer assume most writing is honest or well-meaning because amoral LLMs fundamentally are not wired to make that distinction of true and untrue or right and wrong (unlike most humans) and many people will use and trust what they generate without question, polluting the online space and training data until everything is just a morass of half-known facts sprinkled with generated falsehoods that are repeated so often they seem true.

How do we check sources when the sources themselves were generated by LLMs?

RLHF = Right Left Hand Foot. It's a technique in Bavarian interpretative folk dance where you jump around, artfully hitting the soles of your feet with your hands in order to court women who are busy carrying unbelievable numbers of beer Steins into the mountains.

That's what came to mind when I saw the abbreviation. Then I looked it up:

Reinforcement Learning from Human Feedback.