Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Could the next Leo Tolstoy or Jane Austen be a well-engineered AI software programme? It’s a question that is becoming increasingly pressing as machine language-learning software continues to evolve. No one likes to face their own possible obsolescence — especially not writers, who prefer to believe that literary talent is unique and irreplaceable.
Much of this is just nerves. Today’s AI creative writing programmes are not yet at a stage of development where they pose a serious threat to Colleen Hoover, say — or Charles Dickens. But while attention continues to focus on the possibility of a blanket takeover of human literature by AI, far less consideration has been given to the impending prospect of collaboration between humans and AI. It’s a scenario that is — depending on your point of view — either already here or hovering just around the corner.
Earlier this month, American sci-fi writer Ken Liu, who has a clutch of Hugo and Nebula awards to his name, joined 12 other professional authors for a writing workshop on Google’s Wordcraft. This AI tool, which is based on LaMDA, a non-sentient language-learning model, is not yet publicly available but is billed as an “AI-powered text editor” that can, when given the right prompts from the writer, chip in with helpful descriptions, create lists of objects or emotional states, and, at a pinch, brainstorm ideas.
If you’ve ever accepted a Word doc’s spelling suggestion, you’ve already dipped a toe into the world of machine writing. What is new, is the generational shift in the power of these AI models, which include Google’s Wordcraft and OpenAI’s GPT-3, an artificial neural network with more than 175bn parameters that is being trained to use prompts from humans to create its own poems and works of fiction.
When GPT-3 was launched in 2020, it seemed more like a charming experiment than a serious contender for Machine Most Likely To Replace Shakespeare: the literature it produced felt quaint and slightly off-timbre. But its access to human texts has since broadened. In March, Owain Evans, a research associate on machine learning at the University of Oxford, posted a few poems generated by InstructGPT, the latest GPT-3 variant, on his Twitter feed. The poem below, composed on the subject of Twitter in the style of TS Eliot, indicates that learning and replicating the signature voice of a poet may not be as hard as it once seemed:
Upon the platform of the efficacious tweets
We stand together, faceless and bereft
Of all identity but that which is supplied
By the retweet button and the follow.
We are the retweeted and the followed,
The anonymous and the nameless,
The content-less and the formless,
The bereft of all significance.
Despite the tell-tale signs of a machine hard at work in unfamiliar terrain — “the efficacious tweets”, “the bereft” — the cadence and the gloom are both passably Eliotian. I can imagine this poem passing the Turing test (has a text been created by a human or a machine?) quite easily.
The writers at the Wordcraft workshop, however, emerged with mixed reports. After a few weeks, Liu figured out how to tame his AI dragon: “By taking the seed from LaMDA and saying, ‘Yes, and . . . ’ I can force myself to go down routes I wasn’t thinking of exploring and make new discoveries.” Other writers swiftly discovered its limits. “Here is the problem,” Robin Sloan wrote, “Wordcraft is too SENSIBLE. Which of course is a great success for the language model: it knows what’s sensible! Wow! But ‘sensible’ is another word for predictable; cliched; boring. My intention here is to produce something unexpected.”
I’m unconvinced that writers such as Olga Tokarczuk or Kazuo Ishiguro have much to fear from present-day AI language generating models and writing assistants. Their work, and that of countless other novelists, short story writers, dramatists and poets, is too particular, too beautifully unique; even if a model learnt what they had done in the past, it would not be able to predict where their creative impulses might take them in the future. But for authors who write to a formula — high-voltage thrillers or romances that adhere to a simple template — AI models might plausibly step in, first as assistants before some day making the dazzling leap to authorship.
Assembly-line novels are nothing new. In the 1970s, Barbara Cartland, who wrote more than 723 books in her lifetime, many of them romance bestsellers, would dictate her novels for her secretary Jean Smith to type up at the remarkable rate of roughly seven chapters a week. But already word processors and dictation software have replaced Smith’s role; perhaps creative writing software isn’t that far from replacing the Mrs Cartlands of today.
Join our online book group on Facebook at FT Books Café