[Gormo]

> I have no idea, but the fact is they do!

Well, that's the assumption I'm challenging. That does not seem to be a substantiated fact at all.

Most of the arguments I've encountered online that attempt to make the case that LLMs do have a writing style that's readily distinguishable from human writing are deeply afflicted with confirmation bias, and are often based on circular reasoning.

My comment above was a rhetorical question intended to point out the fact that LLMs are specifically designed to mimic the patterns of writing found in their training data, so the idea that they'd all converge to some other set of pattern is not by itself plausible.

[IshKebab]

> Well, that's the assumption I'm challenging. That does not seem to be a substantiated fact at all.

It's not a assumption or confirmation bias. It's plain as day if you read any of these AI generated (or assisted) articles. You may as well ask me for proof that HN readers are pedantic.

[Gormo]

It's not possible for any of this to be "plain as day" without being substantiated by data. There is no way to validate a statistical indicator without being able to independently measure the target variable.

Validating your criteria against the same subjective impressions that informed those criteria in the first place is an exercise in both circular reasoning and confirmation bias.