There’s no lack of irony in using artificial intelligence to write a novel and calling it Death Of An Author, especially when the man behind the bot doesn’t believe the title is in any way prescient.
Of course, when he isn’t coaxing novels out of ChatGPT, Stephen Marche is writing them himself, so has a vested interest in AI not assuming the literary mantle so far exclusively claimed by humans.
It is an experiment, and one with impressive results. Called ‘the first halfway readable AI novel’ by the New York Times – far from the paper’s most scathing review – it is indeed an easy read, spinning the tale of Canadian writer Peggy Firmin, whose murder literary critic Gus Dupin aims to solve.
No spoilers, but there’s a large AI firm in it too.
They say write what you know, but it is the words, not the story, that have artificial origins.
‘I knew that AI made the most banal plots, terrible plots,’ says Mr Marche, who has written extensively about artificial intelligence. ‘I wasn’t going to use that. I had a conscious process of thinking it through and taking copious notes and figuring out the narrative structure – and from there, using the machines to build it.’
Full disclosure, Death Of An Author is not Mr Marche’s first attempt to write using AI. Described as being 95% by AI and 5% by Mr Marche, it is the former that gets most credit, published under the pen name Aidan Marchine.
‘I’ve been using it to write fiction since about 2017,’ he says. ‘I wrote a short story for Wired called Twinkle Twinkle, which I had a computer scientist create an algorithmic mode of composition for called SciFiQ. Then I wrote another for the MIT Technology Review, and in 2020 I wrote a short story for the LA Review Of Books which was 17% computer generated.
‘Then about a year later I wrote Autotuned Love Story, which was built out of chatbots for Lithub.’
If it feels like Mr Marche is hastening his own redundancy, and that of writers across the globe, a closer look at the process provides some reassurance, if only temporarily. It is a complex one, not yet feasible without a human at the helm.
‘Primarily I used ChatGPT with Sudowrite, the bulk of the text was created using those two technologies in tandem,’ says Mr Marche.
‘I gave ChatGPT very specific instructions, including about grammar, syntax, variability of sentence length, tone, those sorts of things, then followed with “containing the following information”.
‘That would then create a text, which I’d cut, put into Sudowrite, where you can select text and have it shorten it or add detail. It also has a customise button where you can tell it to do whatever you want, such as make it more active, make it more conversational, make it sound like Ernest Hemingway – which I did a lot.
‘I used lots of styles of writers I wanted to imitate, so would have it do that. That was the bulk of the text.’
That may also be the bulk of the argument. Until we reach a point of artificial general intelligence, without human input, chatbots can only create content based on what has come before it.
True, that is literally Everything On The Internet, which makes imitating authors, styles, stories easy for them to do, but at present, without a guiding hand, the raw results are not ready for the real world. Not least because of a belaboured overindulgence of decorative, extravagant identifiers and adnouns.
In other words, overly descriptive and slightly try-hard sentences. Think that episode of Friends where Joey gets a thesaurus.
‘The stuff ChatGPT writes is pretty banal,’ says Mr Marche. ‘To get it to any state where it might be original you need to essentially filter it through a series of stylists and guides, which Sudowrite is.
‘Then when I wanted really good individual lines, for metaphors and things like that, I used Cohere.
‘In that I would create prompts like ‘make an image of coffee’, then I would train that on a series of metaphors for coffee, and it would give me images and I would cycle through them until I found one I liked.’
If this doesn’t sound like the most hassle-free or efficient way to write, know that all of these programs will get better. But, Mr Marche says, the experiment was never about efficiency – even if he was able to produce the novella in a matter of months.
‘Certainly writing with AI is not an efficiency question,’ he says. ‘It doesn’t necessarily make it easier.’
Yet should he and other writers be concerned that it was his agent who pitched it? More likely they spotted a niche idea and timed it perfectly, working in tandem with an author wholly qualified for the project.
However, what AI novels mean for future generations of readers is still to be seen – even if the technology is not entirely to blame. English and literature subjects are plummeting in popularity at universities across the UK – a harbinger of doom for the next generation of authors and editors as they see an industry destined to be razed by technology?
Or will AI enable more people to become authors, finally getting that novel inside of them down on the page – or screen?
‘In this creative form, I’m not sure the term author is appropriate to what I did with this material,’ says Mr Marche. ‘On the one hand, I’m the creator of a literary artifact, so we would obviously call that an author.
But on the other hand, I didn’t create the words, and that’s another way we think of an author that’s probably meaningless when you’re dealing with this text.
‘But the idea that this is going to replace creativity is just not true.
‘This requires creativity, requires human intentionality. It’s really an incredibly powerful, wonderful tool, but whether it leads to the death of the author – I think it’s really more the transmutation of the author.
‘It all becomes authorship, and literary creativity will take on a new form.’
Death Of An Author is published by Pushkin. Available as an audiobook and ebook
