GPT-4 can infer authorship from a passage of text based on style and content

I remember reading about algorithms that could do this years ago (when developed specifically for this purpose). If I'm not mistaken, the plagiarism-detection companies used by educators have pretty reliable software that does this (since most students aren't dumb enough to copy things word-for-word).

Was the blog in the training data?

The training data for GPT-3/4 cuts off in 2021, blog post is from 2023.

People use words and phrases in a distinctive way and the kind of modelling that GPT-4 is based on ought to be pretty good at identifying that (better than it is at real understanding or doing math.)

My prediction: if you picked out 20 pieces of text written by similarly prolific bloggers, it will get it right somewhere between 14-18 cases. 20 is a good number to use for an experiment, 1 doesn't really prove anything. With 20 I am pretty certain you'll see it gets it right sometimes and gets it wrong sometimes.

He should share the entire chat.