@emilymbender I don't have an article, but to summarize, it's pretty much an advanced form of lossy compression for text, running on someone else's server, which one can interface with a web interface that's written in JavaScript which is under a proprietary license.

"OpenAI" has scraped a bunch of text from the internet, most under proprietary terms, some under free terms (but with certain requirements like attribution) and shoved the whole lot through a neural network and the end result is a very energy-hungry language model that accepts question text and vomits out convincing combinations of the input text that matches the question text somewhat, with the output filtered through grammar, spelling and tone checking so it looks very convincing.

By pure chance it gets some things right, but it gets most things wrong due to how much of the input was incorrect and the semi-lossy reproduction nature.