Short answer: em-dashes are used in a lot of professional writing. It's picked up in training, and spread: https://en.wikipedia.org/wiki/Dash#Usage_2

Long answers... I don't really know?

But first some background. LLMs are trained in many stages, with the first being raw text. These 'first stage' LLMs don't behave like chatbots you know; they take a raw passage of text and simply try to complete it word by word, with nothing both the text's context to go on. They're (generally) unbiased, as they've simply been trained on billions of pieces of work with no preference for any particular one.

The subsequent training (to get them to behave more like question/response chatbots) is where they pick up 'style' and quirks, and where trainers may use smaller, more deliberate datasets, and this is where they (specifically OpenAI models, which have this habit in particular) likely pick up em-dashes from. Maybe it's neatly formatted textbooks or summaries, maybe it's reformatting (kind of like how your phone will turn neutral ' " ' quotes into left/right ones), I don't know... But it's there in that stage of training.

They also do what's called "RLHF" at this stage, where they 'pick' answers real humans prefer, and it might be possible that the human testers tend to prefer the neatly formatted em-dash answers more frequently. And all this creates a feedback loop where the more an LLM outputs an em-dash, the more it 'learns' to do it.

...There's also another factor.

It's an open secret that LLMs train on each other's output, deliberately.

As a consequence, they pick up each other's habits. It's possible to even map their "slop patterns" into taxonomical trees, which EQBench attempts to do though its tests:

https://eqbench.com/creative_writing.html

Any notice funny patterns, like how Kimi K2 switched to 'copying' Claude instead of Gemini for its thinking model.

Hence the em-dash habit has spread to other LLMs kinda like slang spreads between human groups.

My guess is, those that do are trained from forum posts where intelligence, including the knowledge of how and the wisdom of when to use non-standard punctuation marks, like en and em dashes, the semicolon, and others, were considered valuable. These people would seem, on the surface, to know more about what they're talking about and would provide better training data for the LLM. Those people used em dashes, so, so too do the AI models based on them.

Also, sorry (not sorry). I am a religious em dash user and have been for over 30 years. I'm not saying I'm smarter than anyone about any one thing, but it is entirely possible some of my forum posts were used to train LLMs. I didn't get paid for it though; hence the "not sorry" part. If it trained on my posts after the fact, I won't take any blame for that. But, people were using em dashes long before AIs were.

State-of-the-art models rely on late-1800s and early-1900s print books for high-quality training data, and those books use ~30% more em-dashes than contemporary English prose. That’s why it’s so hard to get models to stop using em-dashes: because they learned English from texts that were full of them.

That sounds really plausible -- I associate the em-dash with old books and stilted prose, like Sherlock Holmes stories

Theft. Its not even a secret.

This is a cool article.

But if they want LLMs to use fewer em dashes, why not find and replace with a comma or semicolon using a regex that matches known patterns so as to reduce it's frequency in the training data?

If you—like me—had no idea what “RLHF” means and could not figure it from context, it is evidently “reinforcement learning from human feedback”.

Guess they trained on my posts. I’ve always used them.

I think the better question is, why do so many people not use them?

Cause they are EMphatic.