As someone who sits at the intersection of technology and the creative arts, I have a unique perspective on the debate over generative artificial intelligence, or GenAI. It’s a technology that’s been broadly embraced, heavily invested in, harshly criticized, and frequently questioned. It’s a technology that’s been used to summarize articles, write short stories, create artwork, produce songs, and impersonate voices. And it’s a technology that relies on pre-existing content to develop its models; without that content, GenAI doesn’t work.
A controversial part of GenAI has been the collection and use of content without the author’s consent. As a writer, I firmly believe that fair use doctrine doesn’t cover the training of AI models and authors must consent to the use of their work. But I don’t want to use a legal rationale to make that point, because I think there’s a more fundamental argument to be made.
What I want to explore instead is written creative content created by GenAI – specifically fiction – and the fear that it could supplant human creators. That hinges on two criteria: first, that GenAI can create content that matches the quality of human writers; and second, that GenAI can create content that is new, imaginative, and truly original. I want to delve into the latter.
In the literary world, there’s a joke that there are no original stories anymore – everything is just a variation on five or six or eight basic story concepts. And while there is a nugget of truth to that, it’s an oversimplification and, in my opinion, unfair to fiction writers. Some genres and stories may be more formulaic than others, but that’s painting with a broad brush. There are great writers who write great stories that are imaginative and original, even if the rough contours align with stories of the past.
Let’s examine science fiction and fantasy as an easy example. Writers invent characters, worlds, and languages that literally (in the literary and physical meanings) don’t exist. This is the antithesis of large language models (LLMs), which are designed to come up with the “correct” answer, interpreted as the most statistically valid response. By training the model on content that is “correct,” the model learns the right associations between prompts and responses.
When we ask a generic chatbot “You are a greeting card writer, write the text for a happy birthday card,” we expect responses like “Happy birthday!” or “It’s time to celebrate!” or “How does it feel to be one year older?” based on the countless happy birthday greetings the model has been trained on. If it responded with “Congratulations on your birthday!” we would scratch our head; a response of “So sorry for your loss on your birthday” would get it fired immediately from Hallmark.
But if we ask the same chatbot to come up with a plot for a fantasy novel with a heroine who must lead a clan of sorcerers against malicious spirits, what is the “correct” answer? For an LLM, it’s the next token – a word or phrase – that closely matches the previous tokens, based on the original input. So after a flurry of tokens, we end up with the chatbot’s response:
Me: “You are a fiction writer specializing in fantasy novels. Write the plot for a fantasy novel where the heroine must lead a clan of sorcerers against malicious spirits. Your response cannot exceed six sentences.”
Gemini: “The realm of Eloria, once peaceful, is now plagued by malevolent spirits. Elara, a young sorceress with a hidden power, emerges as the only hope. As the spirits grow stronger, she must rally her clan, a group of outcast sorcerers, to protect their home. With courage and determination, Elara embarks on a perilous quest, facing trials that test her limits and reveal the true extent of her power. Her journey will not only decide the fate of Eloria but also unravel the secrets of her own origins.”
Ignoring that Gemini didn’t use its full allocation of sentences – as I’ve demonstrated in a prior post, LLMs can’t do math and fail at the concept of counting – this is a perfectly sensible response by GenAI standards. As a writer, it’s about as exciting as a math textbook. It feels flat and sounds like a generic fantasy story that we’ve read before. Why? Because the model was likely trained on lots of fantasy stories, and it’s stringing together words that it saw most often based on the words in the prompt.
So let’s scold the LLM.
Me: “Make the plot more unique”
Gemini: “In the secluded realm of Eldoria, where magic flows like water, a young sorceress named Elara discovers a sinister truth: the ancient barrier protecting their world is weakening, inviting malevolent spirits to invade. As the spirits begin to sow chaos and despair, Elara, with her unique ability to manipulate shadows, must lead a ragtag group of outcast sorcerers to restore the barrier. Their journey takes them to the heart of the Shadow Realm, where they must confront not only the spirits but also the dark secrets of their own past.”
Great, it picked up the thesaurus and gave us flowery language. But the plot still feels stereotypical and engages in longstanding literary tropes. Let’s scold the LLM again.
Me: “Make it less stereotypical”
Gemini: “In the once-idyllic realm of Eldoria, a young woman named Elara, raised in a secluded monastery, discovers she possesses a unique magical affinity. As the world around her begins to unravel, plagued by increasingly violent spirit attacks, Elara is thrust into a leadership role. She must unite a diverse group of individuals, each with their own unique abilities and personal struggles, to confront the growing threat. Elara’s journey is one of self-discovery, as she learns to harness her power and navigate the complex political landscape of Eldoria, while also grappling with the weight of her newfound destiny.”
As a writer, I’m not impressed. There’s nothing inherently creative about the LLM’s proposed plot. How many times have we seen this in a fantasy story:
- The heroine rises from obscurity
- The heroine discovers her powers, not having previously known they existed
- A peaceful world that devolves into chaos
- The heroine must confront her own personal demons
- The heroine must bring together a diverse group of people in a “stronger together” mentality
- Political drama
This is why I’m convinced that creative writers will prevail over generative AI. There will always be a market for novels and stories that follow a well-worn path, and perhaps GenAI will become a tool in the toolbox for those writers. But what’s exciting in the realm of fiction is a new take, a new angle, a twist on something old. It’s deviating from the stories we’ve read, the characters we’ve followed, the worlds we’ve explored to discover something we haven’t thought of before. And that is something that we won’t get from GenAI.
So how would I rework Gemini’s plot?
“Elara is a powerful sorceress and third-in-command of her clan, which has long struggled to keep the realm safe from malevolent spirits. When a new clan emerges and begins to succeed where Elara’s has failed, she questions whether her clan’s leader has ulterior motives and has intentionally let the spirits prey on the people of Eldoria to help them maintain power. As Elara’s clan quickly loses influence and control, she must decide whether to back her people and risk a war by overthrowing her leader, or switch sides and join the new clan, which may not have her best interests at heart.”
Sounds like something I should consider writing…