I felt personally attacked when LLMs came out: I'm an avid user of "—", bullets, numbered lists, and the word "delve". It's been a miserable couple of years.

LLMs write like that because people wrote like that. Enough, unfortunately for my remaining love of humanity, to cause the LLMs to adopt the quirk.

I know. That’s my point.

People talk about LLM writing style like it’s a unique butterfly and humans don’t write that they. But we do. Which is why LLMs do too.

Many many years ago I wrote a book for Apress, and the style guide for that instilled in me a lot of practices that now make my writing feel LLM-ish to some readers:

- Use bulleted lists, but always introduce and conclude a list with prose; a list can't immediately follow a heading or end a section.

- Use a mix of long and short sentences; in long sentences with parentheticals, use a mix of commas, parens, semicolons, and em dashes.

- With multiple continuous blocks of prose that aren't naturally broken up by an illustration or heading, start a paragraph with an inline bold statement to help anchor the reader.

I still use bullets extensively. You can easily tell when a human writes them when they are trees instead of lists.

I don't think even that is a reliable indicator because I'm currently reviewing an LLM generated bullet tree right now.

Oh interesting. Before the LLM craze, I only ever saw good bullet trees in legalese and git commit messages. The trainwrecks were far more likely to be the rare attempt by HR in a big email or in the odd Jira epic by a PM.

I wouldn't think LLMs would have much to train on. I still see some bad ones, but I don't feel like the quality ratio or overall quantity has changed. I do see more bullet lists though.

For sure, but I don't think I'm going to give Vercel benefit of the doubt that they aren't writing their copy with an LLM.