When Robin Sloan began writing his novel Moonbound, he built a custom language model trained on his own sentences — not to have the machine write the book, but to have it resist him.
He wanted a collaborator that knew his rhythms well enough to push back. "The model would complete my sentences in ways that were almost right," he wrote in a 2023 essay, "and the wrongness was generative." He accepted perhaps one in twenty suggestions. The rest he rejected — but rejecting them clarified what he actually wanted.
The most common mistake writers make with AI is treating it as a faster keyboard. They describe what they want, accept what the model produces, and move on. This is output extraction, not collaboration — and it tends to produce prose that feels hollow precisely because no genuine friction occurred.
Authentic co-writing involves iterative exchange. The writer proposes something imperfect; the AI responds, often inaccurately; the writer revises in response to that inaccuracy; the loop repeats. The value lives in the loop, not in any single output.
In 2022, the author K Allado-McDowell published Pharmako-AI — a book created entirely through alternating passages between a human and GPT-3. Allado-McDowell described the process as "a kind of séance," in which the AI's unexpected associations pulled the human writer into territory they wouldn't have reached alone. The book was published by a literary press and reviewed in the London Review of Books. It is a real collaborative artifact, not a novelty.
Practitioners have settled into roughly three working models, each with different implications for creative ownership:
The AI generates large volumes of raw material. The human curates, selects, and shapes. The author's intelligence is in the editing eye. Ross Goodwin's 2018 road novel 1 the Road — the first AI-authored book published commercially — worked this way: a neural net produced the text; Goodwin shaped the journey.
The human writes primary drafts; the AI responds with alternatives, objections, and extensions. Sloan's approach fits here. The human accepts almost nothing but uses rejection to sharpen their own choices.
Human and AI alternate passages, chapters, or voices. Allado-McDowell's Pharmako-AI exemplifies this. The final voice is genuinely plural — identifiable neither as purely human nor purely machine.
The AI produces complete first drafts which the human substantially rewrites. This is the most common commercial use. The human's voice dominates the final product; the AI's contribution is speed and reduced friction at the blank-page stage.
Every co-writing partnership must answer one practical question: who is responsible for what? Current AI systems excel at surface fluency, genre conventions, structural scaffolding, and variation generation. They are consistently weak at sustained causal logic across long documents, authentic emotional specificity, and anything requiring lived embodied experience.
Experienced co-writers exploit this asymmetry deliberately. They use the AI to draft the connective tissue — transitions, scene-setting, expository passages — while writing by hand everything that depends on felt knowledge: grief, shame, the precise texture of a specific place.
Romance novelist Jennifer Lepp (writing as Leanne Leeds) publicly documented her process of co-writing with Claude in 2023. She found the AI reliable for plot-chapter outlines and dialogue scaffolding, but returned to solo drafting for any scene requiring genuine emotional subtext. Her series output tripled; her reviews remained stable. The partition was deliberate and consistent.
The quality of AI co-writing depends almost entirely on the quality of the human's engagement with what the AI produces — including its failures. Writers who treat every wrong response as information tend to produce better work than those who simply retry until the output is acceptable.
You'll practice the Sparring Partner model. Write a short passage — a scene opening, a character description, a story premise — then ask the AI to continue or respond. Your goal is not to accept the AI's output, but to use whatever it produces (including wrong turns) to sharpen your own revision.
After at least three exchanges, describe one way the AI's imperfect response helped you clarify what you actually wanted.
In 2023, writing instructor Lincoln Michel published an essay in Esquire noting that stories generated by large language models exhibited a characteristic failure: every character spoke in the same slightly elevated, slightly helpful register. The villain explained their motives with the same courtesy as the hero. The grieving mother used complete sentences.
Michel called it "the voice collapse" — and identified it as the central challenge for any writer trying to use AI for character-driven fiction.
Language models are trained to produce text that is coherent, clear, and unlikely to confuse. These are precisely the qualities that make for bad dialogue. Real characters interrupt, deflect, contradict themselves, speak in incomplete thoughts. They have verbal tics, class markers, regional syntax. A model optimized for clarity will sand all of this away unless you specifically instruct otherwise.
The solution is not simply to tell the model "write distinctive voices." That instruction is too abstract. Effective character-voice prompting operates at the sentence level. You must give the model specific examples of how the character sounds, not descriptions of personality.
Practitioners who successfully maintain character voice across long AI collaborations typically provide what might be called a voice specification — a compact block of information the model can reference. It contains four elements:
The AI writing tool Sudowrite, used by a reported 30,000+ fiction writers as of 2024, built its "character voice" feature around exactly this structure — asking users for sample dialogue before generating in-voice text. Internal user research found that providing even two sample lines reduced voice-collapse complaints by a measurable margin compared to personality-description prompts alone.
A single session can maintain voice reasonably well. The problem is persistence: language models have context windows, and as a document grows, early voice specifications drift out of active context. Writers working on novel-length projects with AI collaboration have developed several practical responses:
A short document — 150–300 words per major character — kept outside the main draft and re-injected at the start of each new writing session as a system prompt or preamble. Treated like a style guide for the collaboration.
Every 2,000–3,000 words, the writer pastes a recent AI-generated dialogue passage and asks: "Does this match [character]'s voice card? What has drifted?" The AI identifies its own deviations.
Telling an AI a character is "gruff" or "witty" produces the AI's generic version of gruff or witty. Showing it three lines of actual dialogue — especially including what the character would never say — produces something far closer to a specific, irreplaceable voice.
Build a Voice Specification for a character you're working on (or invent one). Include: (1) three sample dialogue lines, (2) one syntactic rule, (3) one thing they'd never say. Then ask the AI to write a new line in that voice — and evaluate whether it succeeded or collapsed.
Run at least one Drift Check: after getting a response, ask "Does this match the voice spec I gave you? What drifted?"
In 2023, fantasy author Brandon Sanderson discussed AI writing tools in a podcast interview, noting that the fundamental problem he saw with using them for his Cosmere universe was world-state memory: a model given his books as context would still, within a single session, contradict established facts about his magic systems.
He wasn't alone. A 2023 survey of genre fiction writers using AI tools (conducted by the Science Fiction and Fantasy Writers Association) found that continuity inconsistency — characters knowing things they shouldn't, geography changing between scenes, established rules violated — was the most frequently cited frustration, cited by 71% of respondents.
Language models do not maintain a state machine for fictional worlds. They have no internal model of what is true in this story separate from the text currently in their context window. If the text establishing that the city of Valdren was destroyed three chapters ago has scrolled out of the context window, the model will cheerfully have a character visit it.
This is not a bug that will be patched; it reflects something fundamental about how sequence-predicting models work. They are not simulating a world — they are predicting what text is likely given the text they can see. World-state management is therefore always the writer's responsibility.
Professional writers working with AI on long-form fiction have converged on a practice borrowed from television production: the world bible. In TV, a bible is the document writers' rooms use to maintain consistency across seasons and writers. For AI collaboration, it is a structured external document injected into each session.
An effective AI world bible has three layers:
Fantasy writer Aedan Peterson, who documented his AI co-writing process in a series of posts on the Substack "AI Fiction Lab" in 2023, maintained a 4,000-word world bible that he pasted into every session as a system preamble. He reported that continuity errors dropped from roughly 3–4 per session to under 1 per session after adopting the practice. He also introduced a session-end ritual: asking the AI to list every new fact introduced in that session so he could update the bible immediately.
Even with a world bible, errors occur. Experienced collaborative writers use two additional techniques:
After generating a scene, ask: "Review this passage against the world facts I've given you. List any contradictions with established geography, timeline, or character knowledge." The AI is good at catching its own contradictions when explicitly asked to look for them.
Before writing a scene involving revelation or discovery, ask: "At this point in the story, what does [character] know about [topic]? What don't they know?" This prevents characters acting on information they shouldn't have — one of the most common continuity failures.
An AI cannot remember what it cannot see. Every serious collaborative fiction project needs an external document — a world bible — that is re-injected each session. Managing that document, not the AI's memory, is what maintains continuity.
Build a minimal world bible for a fictional world (real project or invented). Include: (1) three Hard Facts that cannot change, (2) a Current World-State with two or three mutable facts, (3) one Open Question you want to stay unresolved.
Then ask the AI to generate a scene set in that world. After reading the scene, run a Contradiction Audit: "Check this scene against the world bible I gave you. List any contradictions."
The U.S. Copyright Office issued a landmark decision in the case of Zarya of the Dawn — a graphic novel by Kristina Kashtanova that had been granted copyright, then partially revoked.
Kashtanova had written the text and arranged the images. But the images had been generated by Midjourney. The Copyright Office ruled that the text and arrangement were copyrightable — because they reflected human creative choices — but the AI-generated images themselves were not, because they lacked human authorship. The line was not "did a human touch this?" but "did a human make the expressive choices?"
As of 2024, U.S. copyright law does not protect AI-generated content as such. What can be protected is the human's selection, arrangement, and modification of AI output. The more a human author has shaped, chosen, and transformed the AI's raw output, the stronger the copyright claim on the resulting work.
This has practical implications: a writer who accepts AI draft paragraphs verbatim and publishes them has a weaker copyright position than one who substantially rewrites them. The Zarya decision made this concrete. Kashtanova retained copyright on everything she made expressive choices about; she lost it on everything she simply generated and accepted.
Getty Images filed suit against Stability AI in 2023, arguing that Stable Diffusion was trained on Getty's copyrighted images without license. The case raised the separate but related question of what rights AI companies have in training data — distinct from what rights human users have in AI output. As of mid-2024 the case remained ongoing. It is one of several that will shape the legal context for all AI-collaborative creative work.
Copyright is a legal question. Disclosure is an ethical one — and the creative writing community has not resolved it. Three positions have emerged among professional writers:
Disclose AI assistance specifically — which elements, which tools, what percentage of word count. Argued for on the grounds of reader trust and market honesty. Allado-McDowell's Pharmako-AI takes this position; the AI co-authorship is part of the book's explicit identity.
AI is a writing tool like spell-check or Scrivener — disclose only what you disclose about other tools (typically nothing). Argued for by writers who see AI as acceleration, not authorship. Jennifer Lepp's public discussion of her process was voluntary, not mandatory.
Disclosure standards differ by genre. Literary fiction has higher reader expectations of pure human authorship than genre fiction. Journalism requires disclosure; romance may not. Several publishing houses issued explicit AI disclosure policies in 2023–2024.
Since 2023, many platforms have imposed their own rules regardless of author preference. Amazon Kindle Direct Publishing, Clarkesworld, The New Yorker, and others have issued distinct AI-content policies ranging from prohibition to mandatory labeling to no restriction.
The legal and disclosure questions are external. There is also an internal question: what kind of writer do you want to be, and what does your process need to feel like your own?
The experience of multiple documented co-writing practitioners suggests a consistent pattern: writers who are explicit with themselves about why they are using AI at each step — not just accepting any output that seems close enough — tend to report higher satisfaction with the work and less anxiety about attribution. The question "did I make the expressive choice here?" is not just legally relevant; it is creatively clarifying.
From the Zarya decision, a practical heuristic: the expressive choices are yours; the generated output, unmodified, is not. Apply this not just legally but creatively. The places where you rewrote, selected, and transformed are the places where you are genuinely the author. Maximizing those moments is both better legal practice and better creative practice.
Generate a passage collaboratively with the AI — a scene, a character monologue, a story opening. Then conduct an Authorship Audit: identify every sentence or phrase where you made an expressive choice (wrote it yourself, or substantially modified AI output), and every sentence where you accepted AI output without change.
Ask the AI to help you map this — it can help identify which parts came from its suggestions. Then discuss: what would strengthening your authorship of the weaker sections look like?