Romancing The Robot

A romance novel ghostwriter ponders what would be lost if she used generative AI for her projects.

 

This is an installment of Automate Me, an ongoing Study Hall series about how AI technology is transforming the media and publishing landscape.

When the hubbub around ChatGPT began, my first thought was how great it would be to outsource my sex scenes. After writing romance novels at a breakneck pace for years, you start to run out of positions.

I’ve been a fiction ghostwriter for almost four years, and have completed almost 30 works in that time. I’ve churned out 50 shades of smut — slice-of-life novellas, multi-book series about motorcyclists, fantasy tomes about werewolves with convoluted political systems — as well as young adult fantasy, new adult romance, and literary fiction, sans smut. I work with authors of all experience levels, from non-professionals crafting their debut drafts, to established writers with dozens of books under their belt. I’ve found a particular niche working with romance authors. Their problem is often not that they lack writing skills, but that they lack the necessary time. 

Maintaining steady sales requires consistent, regular publishing at a pace a single person often can’t keep up. The romance authors I work with self-publish (or “indie” publish) ebooks on Kindle Direct, which sell thanks to reader loyalty and a cascade of marketing tactics that either work with or beat the Amazon recommendation algorithm. Once an author establishes their brand, their readers become so loyal and voracious that authors can’t possibly keep up with how much their readers consume. (Some fans read 7-10 books per week, one author tells me.) That’s where ghostwriters like me come in. To keep up with demand, authors outline the narratives of their novels, providing character information and chapter-by-chapter breakdowns, then hand off the actual manuscript drafting to me. I match tone, voice, and revise the stories in collaboration with the author: they’re the conductor, and I’m the first chair. 

As a genre, romance has expanded far beyond the Fabio-covered bodice rippers of yore. Romance has hit the mainstream with the massive success of authors like Colleen Hoover, whose self-published success led to a traditional publishing deal. The genre generates $1.3 billion in annual sales in the US alone, and that number doesn’t include the millions of self-published ebooks — which offer writers a way to make money outside of traditional publishing structures — that are sold each year. If you can develop a dedicated audience, you can make serious money. (Authors can earn up to 70 percent royalties from Amazon ebook sales, delivered monthly.)

Authors are in control of the entire publishing process, from drafting, to cover design and page layout, to marketing the final product. They may not have the financial resources or backing of a large publisher, but they also aren’t beholden to a publisher’s rules, nor its slow pace. That freedom drives experimentation and curiosity in both the creation and marketing of a story. Because the industry is so self-driven, the romance genre is a good testing ground for experimenting with AI-assisted storytelling. And many writers are already doing just that.

I considered doing the same, drawn to AI’s promises to break through writer’s block and skim past dull sequences. I was intoxicated by the idea of writing more words in less time, and thus taking on more contracts and making more money. But generative AI feels like a less sexy version of a coke-fueled idea binge — helpful and fun once or twice, but long-term use may end in disaster.

Generative AI, like OpenAI’s chatbot ChatGPT, produces text based on user input. ChatGPT is built on a large language model (LLM), which is trained to break language into statistical patterns using massive amounts of language data from across the internet. GPT-3’s training corpus includes everything from  Wikipedia pages, Reddit posts, and fanfiction, to potentially copyrighted material used without consent. 

In my private conversations with other romance authors, I’ve encountered curiosity and some interest in using generative AI as a tool, versus the outright rejection I’ve seen more widely from journalists and screenwriters. Generative AI tools are particularly beneficial to self-published writers, because they have immediate, tangible benefits with fewer obvious drawbacks. Fiction doesn’t have the same fact-checking requirements newsrooms do, nor do self-published authors have to grapple with the ethical considerations of editors. Machine learning is already embedded in the self-publishing world, as authors are beholden to the inscrutable forces of the recommendation algorithm. I was already writing manuscripts from outlines that were developed based on trends found on the Amazon romance front page, so why not deploy machine learning to assist my actual writing process?

Romance is a particularly formulaic genre that contains repeated tropes and familiar structures — the kinds of patterns AI tools can be trained to identify. For readers, the familiar story structure is comforting, and the guarantee of a happy ending is a balm in an unpredictable world. It goes like this: two characters meet, chemistry sparks, they get together, break up, then choose love and get back together for their Happily Ever After. Unlike other genres like mystery or thrillers, in romance, the emotional arc is the “A” plot, and the external events are the “B” plot. Every plot beat hinges on the characters’ emotional landscapes: what they think they want, what they need, what wounds keep them from accepting love, and their journey to change.

The emotional arc may be familiar, but it also must be convincing. Readers expect coherent “emotional logic” — when the characters make stupid decisions, the reader must understand why even if it frustrates them. They also expect “chemistry” — interactions should crackle with tension, and the sex scenes shouldn’t just be pornographic descriptions of sex-choreography, but a climax (sorry) to a slow build of relationship development. Romance is wish fulfillment fiction, and not just the wish of being swept away by a billionaire or a werewolf. The wound-healing, seeing two damaged people find solace and growth in each other, is part of that fulfillment.

In short, a romance writer must create a story that is familiar, novel, emotionally logical, gripping, satisfying, and of course, titillating. And we carry out this task over weeks or months, not years. 

I was recently face-down on my desk at midnight struggling through one final relationship-consummating love-making marathon. The scene needed to bring my characters together, show their deep emotional bond, and it needed to be hot. I was exhausted. It felt impossible to write, even though it was the kind of scene I’d written dozens of times before. I opened LAIKA, a generative AI tool built on OpenAI’s GPT-2. LAIKA allows users to train a language learning model on their own writing and the resulting LLM is called a “private brain.” According to LAIKA co-founder Charlene Putney, the company is exploring how to monetize “private brains.” In her envisioned model, an LLM trained on a writer’s voice becomes a piece of intellectual property, owned by the writer, which can then be licensed out for other writers to use as a writing tool. So if James Patterson joined this model, an LLM would be trained on all of his works. Then, for a licensing fee, a budding thriller writer could use Patterson’s “private brain” to write “with” him. 

For ghostwriters, it’s a particularly harrowing thought: my craft of matching an author’s voice becomes an AI-assisted, “de-skilled” task. For just $10 a month, I could upload a minimum of 10,000 words of original prose to the software to generate text in my voice and style. My very own private brain. If a generative AI tool was trained on all the smut I had already written, it would be able to replicate some of what my own brain already does. 

Sex is sex. Did the words really need to come totally, organically from me? I could think of the tool as a collaborator, easing my burden the same way I ease those of my author clients. And yet I couldn’t bring myself to do it. The tool wasn’t just trained on my work, but all the writing GPT-2 had absorbed without other writers’ consent. It wasn’t just an ethical concern that stopped me, but one of client-facing honesty: these authors chose to work with me, not me plus all the work inside a generative AI tool. Writing with AI assistance feels like writing half-distracted; it invites me to skip through the challenges of creative writing to write fast, write efficiently, write more more more. But the challenging moments, the face-down-on-desk-at-midnight moments, have sometimes led to my most electric prose. The pain is part of the process.

Generative AI could be a tool of stratification: writers like myself, who rely on volume, use it to increase output, and readers begin to expect more and more novels at a breakneck pace, as a result. On the publishing side, if I suddenly begin turning over drafts of 100 thousand words in four weeks, instead of eight, an author’s expectations will shift accordingly. The quest for efficiency will erode the quality of writing.

Established writers could begin making money without writing a single word. 

When readers are coming to a genre rooted in an experience so deeply human, it feels disrespectful to offer something that’s only halfway human itself. I’ve decided to lean into the human side of writing. I’m moving away from genre ghostwriting and into closer, more long-term collaborations, working with authors one-on-one for months or years to develop projects intended for mainstream publishing. I can hone in on the emotional heart of a narrative, walk an author through every step of story development, and mold my voice into theirs. I no longer advertise my writing speed. Instead, I advertise to authors that I understand them.

Fiction writers will need to understand AI tools, whether they use them or not. I won’t leave romance behind, but I’m shifting into slower markets for my sanity. I might still dabble, but I’ll stick to writing my own sex scenes.

---

COMINGS AND GOINGS

—Stephanie Palazzolo is leaving Insider and is joining The Information to lead their newsletter on developments in AI.
—Chris Panella is starting a new role as news reporter at Insider.
—Susan Fowler Rigetti is now an editor for Slate’s “Future Tense” vertical.
—Kate Wagner is now an architecture correspondent for The Nation.
—Derek Brower is now the US political news editor at Financial Times. 

---

EVERYTHING ELSE

—Last week, National Geographic laid off all of its staff writers and announced that starting in 2024, it will no longer be sold at newsstands.The monthly magazine will still be sent to subscribers. This is the latest round of cuts at Nat Geo, which in April, laid off the audio team responsible for the Overheard podcast. In February, Bob Iger, CEO of Disney, the magazine’s parent company, announced that he’s slashing $5.5 billion from the budget as part of a broader restructuring.
—Google is throwing a temper tantrum because it doesn’t want to cough up some dough. After the Canadian government passed the ​​Online News Act, which requires Google and Meta to pay newsrooms a fee for linked content, Google announced that it will stop news articles from showing up in search results and in their other products. Earlier this month, Meta said that due to the legislation, it’s pulling news from Facebook and Instagram. According to Canada’s Parliamentary Budget Officer, if the tech behemoths complied with the law instead of being petulant brats, Canadian newsrooms would have received $329 million — that’s peanuts compared to how much Meta and Google make.
—Last week, the GMG Union said that G/O Media informed them that it will start testing artificial intelligence to create content on its websites. The union, which is represented by the Writers Guild of American East, said in a statement that it’s “appalled by this news.” They added, “The hard work of journalists cannot be replaced by unreliable AI programs notorious for creating falsehoods and plagiarizing the work of real writers.”