My first thought was watermarking. Same for it's affinity for using emojis in bullet lists.

This has always seemed intuitively obvious to me. I use a lot of em dashes... because I read a lot. Including a lot of older, academic, or more formally written books. And the amount used in AI prose has never struck me as odd for the same reason. (Ditto for semi colons).

The truth is ... most people don't read much. So it's not too surprising they think it looks weird if all they read is posts on the internet, where the average writer has never even learned how to make one on the keyboard.

Delve on the other hand, that shit looks weird. That is waaay over-represented.

"If AI labs wanted to go beyond that, they’d have to go and buy older books, which would probably have more em-dashes."

Actually, they wouldn't have to go and buy these old books: The texts are already available copyright free, due to legislation stating that copyright expires 70 years after the author's death (any book published in the USA before 1923 is also reproducible without adherence to copyright laws), making the full texts of old books much easier to find on the internet!

Very interesting topic. I also wonder why other signs of AI writing, such as negative parallelism ("It's not just X, it's Y"), are preferred by the models.

Also, I wrote a small extension that automatically replaces ChatGPT responses with em dashes with alternative punctuation marks: https://github.com/nckclrk/rm-em-dashes

What we also learned after GPT-3.5 is that, to circumvent the need for new training data, we could simply resort to existing LLMs to generate new, synthetic data. I would not be surprised if the em dash is the product of synthetically generated data (perhaps forced to be present in this data) used for the training of newer models.

I am no grammarian, but I feel like em-dashes are an easy way to tie together two different concepts without rewriting the entire sentence to flow more elegantly. (Not to say that em-dashes are inelegant, I like them a lot myself.)

And so AI models are prone to using them because they require less computation than rewriting a sentence.

As someone who used em-dashes extensively before LLMs I can only hope (?) some of myself is in there. I really liked em-dashes, but now I have to actively avoid them, because many people use them as a marker to recognize text that has been invented by the stochastic machine.

This episode of Big Technology Podcast goes into the reason why:https://pca.st/episode/4090833a-2abd-42b2-a31d-ebb2b4348007

Another reason I think attributes to it at least partially is that other languages use em-dashes. Most people use LLMs in English, but that's not the only language they know and many other languages have pretty specific rules and uses for em-dashes. For example, I see em-dashes regularly in local European newspapers, and I would expect those to be written by a human for most part simply because LLM output is not good enough in smaller languages.

I've been using em-dashes in my own writing for years and it's annoying when I get accused of using AI in my posts. I've since switched to using commas, even though it's not the same.