Must writers now collaborate with AI to stay ahead?

A stylized illustration shows ideas – glowing swirls of gold – flowing from a fountain pen onto a page. It captures the current moment in creative writing, when authors often feed words to an algorithm and watch it fill the page. In effect, a writer can prompt ChatGPT or a similar large-language model (LLM) and receive fresh text on demand. Advocates call this a way to “amplify” human creativity^[1][2]. Critics, however, note that machine-produced prose still lacks the soul of human art – it often falls back on pattern and cliché^[3][4]. In practice, the secret is learning how to talk to the machine: giving it just enough guidance and style so that it can spin out useful story ideas or vivid descriptions without losing the writer’s voice.

AI-generated fiction isn’t entirely new. In fact, by the mid-2010s experiments in creative text were already under way. The first openly AI-written novel appeared in 2018: 1 the Road, composed by a “machine curating road-trip data”^[5]. Two years earlier, a Japanese team had nearly won a literary prize with a short story co-written by an AI^[5], and that same year a short sci-fi screenplay – Sunspring – was produced by a neural network that beat out human writers at a London film festival^[5]. Since then, text-generation tools have leapt ahead. Today’s models can craft pages that seem astonishingly human. One commentator notes that modern GPT-3 engines “can produce language of astonishing clarity and creativity” and even move us emotionally^[2]. In practice, however, the model only “surprised” when it stayed within plausible pattern: a tragic essay written by Vauhini Vara rose to fame in 2021 partly because only the final line was truly inspired by the AI – the rest Vara wrote herself^[6][7]. In other words, AI can be a trigger or collaborator, but not the sole author.

The Probabilistic Muse: What’s Lost and Gained

All LLMs share a fundamental nature: they predict words by crunching statistical patterns in their training data. When a user types a prompt, the model produces what it “thinks” are the most likely next words. As one sci-fi publisher recently lamented, these algorithms are “not built to entertain or surprise us” but merely “analyze patterns” to guess a plausible response^[4]. Because of that, AI writing often tilts conventional. Even Sanjida O’Connell (writing on the Royal Literary Fund blog) warns that, in her experience, “almost everything I’ve read that ChatGPT has written for me so far has been bland and clichéd – helpful only as a starting point”^[3]. Another novelist observed that a computer “could never tell you the way Karenin smiled, nor would it ever fixate on all the place names that filled Proust’s childhood”^[8] – meaning the little human nuances of character and atmosphere tend to be missing from AI output.

Yet the AI also holds some unexpected strengths. Because it has ingested enormous amounts of text, an LLM can suddenly sketch a credible example of almost anything. One writer noted that after giving GPT-3 a modest prompt about her late sister, the result was mostly trite – but it was something to react to. It gave her tangible prose to work against, sparking her own voice. In the final published essay, only one line came from GPT-3^[6]; the rest was filled in by her. Without that machine-provided line, she later admitted, she might have written nothing^[6]. In other cases, users have tweaked the model into serving particular roles. A college student found that telling ChatGPT “act as a [genre] novelist” or explicitly outlining context (setting, characters, word count) produces far better results than a simple “write a story” command. In short, the AI’s co-writing ability depends entirely on smart prompting.

In some domains, the technology has already impressed readers. Gamers who used AI Dungeon – an interactive storytelling game powered by GPT – reported a sharp rise in “creativity and storytelling ability” after an upgrade to GPT-3, calling the effect “heavenly”^[9]. In one anecdote, players felt the AI even adapted to explicit themes with nuance. In others, problems arose: stories veered into bizarre or banned content, and Dungeon’s moderators struggled to filter it. This highlighted a key limit: the AI does not understand context or ethics the way people do. It will happily produce any text that matches its learned patterns, from a magical tale to unsavory fantasy, depending on how it is prompted. OpenAI has even warned that the ease of generating hundreds of thousands of words with these tools challenges our notions of authorship and originality^[10][4].

Engineering the Story: How to Prompt an LLM

Given these quirks, successful writers treat AI prompts like a craft. They do not dump a half-baked idea and hope for a masterpiece. Instead, they design prompts with care, almost like outlining scenes for a ghostwriter. The basic rule: provide clear context, specifics, and style cues. For example, one recommended prompt is literally to ask ChatGPT to “act as a member of a creative writing group” and brainstorm ideas around a premise^[11]. A user might write: “I’m working on a fantasy novel about a world without water. Act as my creative writing partner: for each idea I share, generate five more.” This tells the model exactly what role to play and how to respond. Likewise, if stuck on writing prompts themselves, writers can ask: “I want to write a 500-word flash fiction about loss. Please suggest five creative writing prompts focusing on emotional connection”^[12]. Being specific – in tone, length, and theme – is key. Open-ended requests like “write a story” often yield dull boilerplate, but detailed instructions coax the AI toward useful, varied output.

Another tip: Iterate in conversation. Chat interfaces allow back-and-forth. A writer might start by asking for a list (characters, conflict ideas, taglines, etc.), then pick the best suggestions and ask the model to expand or refine one. For instance, after getting plot ideas, one can say, “Take the third idea and write an opening paragraph. Use a moody, first-person tone.” Because the model remembers the thread, it can continue in style. This mirrors real writing workshops: the human and AI pass drafts and suggestions to each other. O’Connell stresses that her own process with GPT-3 ended up inverted – by the end, the AI was prompting her more than she prompted it^[13]. That is, she gave progressively detailed briefs (“My sister’s cancer was Ewing sarcoma…”), correcting its euphemisms, so that ultimately the algorithm’s weakest line (“she went into remission… she’s doing great”) highlighted how she needed to assert her truth^[7][13].

Voice and style cues can also help. Many prompt-engineering guides suggest starting with a role or genre: “You are a noir detective narrator in 1940s Los Angeles. Continue this scene…” or “In the style of [classic author], describe…” (Note: using a living author’s style is legally iffy, but citing a public domain author is fine.) Setting constraints – for example, “write a 200-word scene” or “use present tense and vivid imagery” – guides the AI and avoids rambling. Conversely, telling the AI not to do something (like “avoid clichés”) can sometimes be risky, since the model might latch on to the prohibition. Better to model good examples. If one wants multiple versions, try chain-of-command: first ask for an outline, then ask for a short story based on that outline, etc.

One should also verify facts. LLMs can “hallucinate” details (invent fake people or plot points). The best practice is to treat AI output as a draft or creative inspiration, not a finished product. As the Royal Literary Fund notes, many authors still hire editors and fact-checkers – because machines sometimes make confident-sounding but incorrect statements^[14][3]. In fiction this might mean strange inconsistencies. It’s on the writer to catch those. Remember: the AI doesn’t truly imagine; it cleverly edits what it has seen. What it gives us is always influenced by human writing already out there. We remain needed to add original flair.

Tales from the Prompt Front: Projects and Pitfalls

Across literature and the web, ambitious projects have put these ideas to test. Vauhini Vara’s “Ghosts” essay (2021) is one of the best-known examples. Vara, a journalist and novelist, used GPT-3 to help her write about her sister’s death. She fed the AI paragraphs from her own life and let it “continue the story.” The result was a patchwork: moments of genuine resonance interspersed with formulaic filler. A sample prompt about her sister’s cancer yielded a collegiate tone – “Eventually, she went into remission… She’s doing great now.” – something Vara quickly abandoned^[7]. But as her New Yorker reviewer notes, GPT-3 did give her a way to begin: in the most gut-wrenching final section, only two lines were AI-generated – the rest was her own raw voice^[13][6]. In practice the AI served to unlock her writing, not write it. She later reflected that GPT-3 “delivered words to a writer who had been at a loss for them,” but only as a springboard for her own truth^[13][6].

In more commercial fiction arenas, AI has stirred controversy. In Japan in 2025 an AI-authored romance novel rocketed to #1 on a big web-fiction site^[15][16]. Its author, “Natsumi Nai”, pumped out the story by continuously prompting an LLM with new chapters – roughly 100,000 characters per day – to exploit the site’s ranking algorithm. Readers discovered the deception and it sparked debate: some hailed it as “creative evolution,” but many writers were alarmed that human-paced storytellers were being outrun by a machine-fed flood^[15][16]. The episode highlighted a problem: AI can churn out quantity at inhuman speed, but critics note the quality of such flooded fiction often feels thin. Similarly, Clarkesworld, a respected sci-fi magazine, was recently inundated with spam of AI-generated short stories. Editor Neil Clarke said his queue jumped from nearly zero AI submissions to hundreds a month – literally thousands of pages written in moments by bots^[10][17]. He pointed out dryly that one human writer might labor for months, while AI could “churn out hundreds of thousands of works” in the same time^[10]. And yet, he found the machine-text “not even any good” – a stream of predictable tropes. The magazine was forced to close submissions temporarily and reconsider how to distinguish human art from algorithmic imitation^[17][4].

Still, it isn’t all doom and gloom. Plenty of writers see AI prompts as tools rather than threats. Some novelists now rely on specialized apps: Sudowrite and NovelAI are platforms built for fiction authors to brainstorm on demand. These tools often package the kind of prompt-engineering advice discussed above into slick interfaces. One indie writer noted on a blog that AI is “great” for simple idea-generation, plot outlines, character sketches, even finishing a chapter from a prompt^[18]. Another example: a tech journalist used GPT-3 to draft snappy taglines and back-cover blurbs for novels – tasks she found harder than writing the story itself^[19]. In marketing and promotion, AI can spit out dozens of headline and blurb options in seconds. (Though here too, writers usually polish the final text to make it sound authentically human.) Some small-publisher experiments have even used AI to illustrate scenes: prompting DALL·E to sketch characters or covers when commissioning a real artist isn’t feasible^[20].

In all these cases, the lesson is similar: AI can amplify certain phases of storytelling, but it isn’t a surrogate for human imagination. The patterns in GPT’s output can sometimes surprise us, but they ultimately reflect the collective human culture it was trained on. A recent piece in The Philosopher put it this way: “AI-generated fiction needs to be grounded in the overlap of reading and human life,” because the “alchemy” of creativity still comes from lived experience^[5][2]. Or as one developer friend of a writer bluntly advised years ago: encouraging a bot to do your thinking is to “strip everything good, original, and beautiful from the creative process”^[8].

Exercises: Your Turn to Try AI Prompts

• Brainstorm story ideas. Pick a genre and premise (e.g. “space opera with an aging starship captain”). Prompt the AI: “Act as a creative writing partner and brainstorm five story ideas for [this premise].” Use the model’s suggestions to spark fresh plots or twists^[11].
• Generate writing prompts. Invert the idea: ask the AI to write prompts for you. For instance: “I want to write a 500-word flash fiction about loss. Please suggest five creative writing prompts focusing on emotional connection”^[12]. Then choose one prompt to actually write your own piece.
• Character backstory. Give the AI a character sketch or name and ask it to flesh out details. E.g.: “Create a biography for Elina Dray, a 30-year-old cryptographer who hears voices.” It might suggest a history or conflict you hadn’t imagined, which you can then incorporate.
• Continue a scene. Write a short paragraph of your own, then copy it into ChatGPT and request “Continue this story for three paragraphs”. Or vice versa: give the AI an opening line and let it build the first scene. Compare its version to what you would write.
• Dialogue writing. Supply a situation and two character descriptions. Prompt: “Write a dialogue of 300 words between a disillusioned robot and a novelist in a cafe at midnight.” This can show you how the AI handles conversation dynamics and voice.
• Genre twist. Take a familiar premise and ask ChatGPT to recast it. For example: “Rewrite the story of Cinderella as a science-fiction tale on Mars.” This tests the AI’s creativity and can generate ideas you might not have thought of on your own.
• Descriptive imagery. Give the AI a setting and mood: “Describe a moonlit London alley, in the style of a Gothic horror.” See if the output sparks visual details you can weave into your story. (Remember: you may need to refine or correct the details it invents.)
• Multiple endings. Write the beginning of a story yourself. Then ask the AI: “Suggest three possible endings for this story.” The AI’s alternatives can reveal interesting directions or moral twists, even if you end up discarding some.
• Stylistic re-write. Take one of your paragraphs and have ChatGPT “rewrite this paragraph in the style of [a well-known author or genre],” being careful not to copy any copyrighted prose. Use its version to learn about tone or phrasing, but remember to blend in your own creative voice.
• Edit and critique. Paste a draft chapter (up to 500 words) and ask for specific feedback: “Proofread this chapter for grammar and style. Then suggest two ways to improve the pacing or character development.” Treat its critique like a friendly editor’s–consider its points but use your judgment.

Each exercise is meant to engage the AI – not replace you. Push and refine its outputs. Ask “why” or “what if” questions to make the prompt iterative. And always remember: no matter how clever the machine, you still hold the imagination that started it all.