Can AI Write Good Romance? What Research and Authors Say

AI can write competent romance sentences. Whether it can write good romance, the kind readers reread and recommend, is a different question, and the honest answer in October 2026 is: not on its own. The evidence we could verify says AI help raises the floor for weaker writers and pulls stories toward the average. Romance readers notice the average faster than almost anyone.

This piece separates what is reported from what is opinion. The first three sections are sourced facts. The craft sections after that are our view as a platform that hosts an AI editor, so weigh them accordingly.

Romance is one of the few print categories still growing. Publishers Weekly, reporting Circana BookScan data in January 2026, said romance unit sales rose 3.9% in 2025 to almost 44 million, while overall print units were roughly flat at 762.4 million and fantasy fell 8.7%. Much of romance’s gain came from Rebecca Yarros’s Onyx Storm.

Those are print units at outlets that report to BookScan; they leave out most ebook and subscription reading, where romance is also strong. But even the partial picture explains the pressure. A genre with huge, fast-reading audiences and a long tradition of series and pen names is exactly where someone will try to scale output with software.

Someone has. According to The Decoder’s summary of a New York Times report from February 2026, a romance author writing as Coral Hart said she produced more than 200 novels in a year with AI models, including Anthropic’s Claude, under 21 pen names, without disclosing it to readers. The same report quotes her own complaints about the output: elegant prose, but weak banter, and chatbots that struggle to build tension and rush to the payoff. Those claims are self-reported and we could not audit them.

The genre’s formal bodies lean hard against AI-written text. The UK’s Romantic Novelists’ Association, in a position statement updated in July 2026, says generative AI “leads to homogenous stories and strips away unique voices,” encourages members to avoid it for writing or editing, and requires anyone entering its awards or New Writers’ Scheme to confirm the work is entirely human-authored, with tools limited to spelling, grammar and research.

Rank-and-file reaction has been just as sharp. In February 2024, 404 Media reported that a Romance Writers of America chapter in Idaho pulled a Zoom workshop promoting AI in romance writing from its site after backlash from authors.

Use is real but narrower than the headlines suggest. In a May 2023 survey of more than 1,700 writers across genres, the Authors Guild found 23% used generative AI at all; of those, 47% used it as a grammar tool, 29% for brainstorming plot and characters, and only about 7% to generate the text of their work. Ninety percent said authors should be paid if their work trains AI. The survey is not romance-specific, and it is now three years old, so treat it as a baseline rather than today’s number. For the reader side of the trust problem, see AI slop books and reader trust.

The most cited experiment is Anil Doshi and Oliver Hauser’s 2024 paper in Science Advances. Participants wrote very short stories, about eight sentences, either alone or with access to one or five GPT-4 story ideas. Stories written with AI ideas were rated more creative, better written and more enjoyable, with the biggest gains among writers who scored lowest on a creativity test. Writers who scored highest gained little or nothing.

The catch was collective. AI-assisted stories were more similar to each other, and to the AI ideas, than stories written by people alone. TechCrunch’s write-up put the shift at roughly 9 to 10% closer to the average: less varied as a group, not clones. The authors call it a social dilemma, where each writer is better off and the shelf as a whole gets narrower.

Eight-sentence stories are not novels, and the study did not test romance. But the direction matches what romance readers say they dislike in AI text: the same beats, the same phrasing, the same safe emotional temperature across books that are supposed to have distinct voices.

Our view: AI is most useful on the parts of the job that are about bookkeeping rather than feeling. It can hold a continuity sheet, so eye color, the name of the bar and the timeline of the first kiss stay consistent across thirty chapters. It can outline, or better, critique your outline and point out where the middle sags.

It is also good at variations. Ask for five ways a scene could end, then write the sixth one yourself; the menu tells you what is predictable, which is useful information in a genre where readers know the tropes as well as you do. Line edits for clarity, rhythm and repeated words are the safest use of all, and the one most writers in the Authors Guild survey reported.

Clichés first. Models are trained to produce the likely next phrase, and in romance the likely phrase has been written ten thousand times. Left alone, they reach for the same breath catching, the same jaw tightening, the same rain at the right moment.

Tension second. Good romance is mostly delay: the thing the characters want and refuse, the conversation that goes sideways. Models tend to resolve conflict quickly and explain feelings out loud, which flattens the slow build. Even the AI-heavy author in the Times report said as much.

Third, agency and consent. A love interest whose choices are clear on the page reads as a person; one who simply reacts to the plot reads as a prop. AI drafts often blur who decided what, and in romance that is not a style issue. Readers want desire that is mutual and legible, between adults who choose each other. A human has to own that.

Finally, voice. The Doshi and Hauser result is about sameness across writers. If many authors lean on the same models for the same beats, their books start to sound alike, and voice is the one thing a romance reader follows from book to book.

Our suggestion is simple. Write the first draft of every emotional scene yourself, even badly. Use AI after, never before: ask it to flag continuity errors, overused words and places where the pacing drags, then decide each change yourself. Use it for structure questions, not for the sentences readers will quote.

Keep a short voice sheet, a few lines on how each lead talks and what they would never say, and check AI suggestions against it. If a suggestion would fit any book, cut it. We walk through this in more detail, scene by scene, in how to write slow-burn romance with an AI editor.

And follow the rules where you publish. Some stores and contests require disclosure or forbid AI-generated text entirely, as the RNA’s statement shows. This is an overview, not legal or contract advice.

ShortFling is a fiction platform, not an AI chat or companion app. Our writing studio includes an AI editor that is free within a daily allowance, meant to help you revise a draft you wrote, not to write it for you. Our creator agreement requires original work and adult characters, and pressing Publish submits a story for review; it goes public only after approval.

If you would rather read than write, the fiction library has general-audience stories and an NSFW section for readers who confirm they are 18+, where availability varies by region and platform. Judge for yourself whether the voices on the page sound like people.