authentic is as authentic does, or is it?
an annoying thought experiment
I have quite a lot of posts on substantial issues in the chamber at the moment, but Friday is for light hearted philosophical squabbling. So, here’s a thought experiment, based on my intuition that a) one kind of literature that LLMs might be very good at churning out might be “misery memoirs” but b) people wouldn’t buy them if they knew that there was no actual misery involved.
Consider three books:
“The One That Made It” is the story of a boy from Yorkshire who overcame a horrific childhood to become one of Britain’s most decorated special forces soldiers. It is 100% AI slop. The author is an Oxbridge graduate who noticed that books like this were selling and got an LLM to generate one. They are themselves a professional writer, and they put quite a lot of effort into the book, but everything in it is 100% artificially generated.
“Two Kinds Of Hardness” is the story of a boy from Yorkshire who overcame a horrific childhood to become one of Britain’s most decorated special forces soldiers. It is 100% genuine. The author is a quite exceptional person, who not only went through the intense experiences described in the book, but turned out to have a lot of literary talent and wrote a captivating memoir.
“Three Thousand Miles From Home” is the story of a boy from Yorkshire who overcame a horrific childhood to become one of Britain’s most decorated special forces soldiers. It is a genuine story, but it is ghost-written. All the experiences described in the book happened to one person, and all the words were written by another person who didn’t have those experiences.
“Four Ways To Live” is the story of a boy from Yorkshire who overcame a horrific childhood to become one of Britain’s most decorated special forces soldiers. The author is a quite exceptional person, but not really any good at writing. So he used an LLM to generate the text. All the experiences described in the book happened to the person whose name is on the cover, all the words in the book were generated by a machine in response to prompts from that person and then approved by that person.
Because this is an annoying philosophy thought experiment, assume that all the actual words in all four of these books are exactly the same, and ignore all questions of whether that makes sense or not.
I think there’s pretty strong consensus that Two is the apex and One is the nadir of the scale of “authenticity”. But what order would you put Three and Four in between them? And where would you place the cut-off in terms of something you’d regard it as acceptable to sell to the public as an autobiography? (Note that there is potentially a continuum of LLM use and ghostwriter involvement – how much help in generating ideas and organising themes could the author of Two get before you start saying that this is more like Four?)
Actually, shortly after pressing publish I've come up with the even worse "Zero Hour", in which someone hears the true story of a Yorkshire boy &c on the radio, realises that it might sell as a book and fires up the LLM. So the experiences happen to a real person, but one that was not involved in creating the book. My intuition is that this is a) much worse than One and b) also worse than doing the same trick but writing it yourself. Am I right? (Because if I am, I think this undermines the kind of theories based on a causal link with reality which I was going to try to use to justify my own rankings)
How is this post not called "ChatGPT, Author of the Quixote"