It works by predicting the most likely words to follow the sequence you already have, with a bit of noise added. The result is that if you ask it a question, it is effectively designed to sound as much like an answer as possible. Whether or not that answer is true is out of scope, and not something that technology could ever consider.

I like to talk about it as if it’s the world’s best prop master. You ask for a prop, and you’ll be given one of the most realistic props imaginable. You want a medical chart, it will give you a chart that might fool a doctor. You ask for a legal brief, it has one that might just fool a judge in court. If you ask it for a computer program, what it spits out might actually compile and/or run. But, of course, these are props. They’re only designed to look good on camera. At most, someone will stream it in 4k, pause it, and try to read the prop while it’s on screen.

As someone who has been annoyed with props for decades, I love that. No more “computer code” scenes where it’s just random gobbledygook. But, someone trying to use the output as if it’s real is just as clever as someone who tries to spend prop money from a movie.

it says as you type that it can make mistakes