Ask GPT to write you hard sci-fi in the vein of The Three-Body Problem, or a Harry Potter-style fantasy, or a pulpy web novel. By around the 10,000-word mark it will start talking gibberish — characters break, plot lines tangle, the whole thing collapses.

Most people quit here and conclude: AI can’t write a novel.

Forget the argument for a second. Look at four facts.

  1. January 2024: Japanese author Rie Kudan won the Akutagawa Prize, Japan’s top literary award, for Sympathy Tower Tokyo. She openly admitted about 5% of the book came directly from ChatGPT. The judges called it “almost flawless.”
  2. American indie sci-fi author Joe Vasicek used to spend 6 months to 2 years on a book. His most recent novel — 18 chapters, 80 scenes, 153,254 words — took 5 days of AI generation.
  3. American genre author Leanne Leeds has 5 series, 40+ books, a 4.5-star Amazon average. Beta readers can’t tell she uses AI. She finishes a book in about 3 weeks.
  4. March 2026: Hachette, one of the Big Five publishers, urgently pulled a horror novel called Shy Girl. It launched in November 2025, blew up on BookTok, and racked up over 4,900 ratings on Goodreads. The reason for the pull: an AI detection company called Pangram measured the book at 78.4% AI-generated.

The tools they use are mostly ones you’ve never heard of. The workflows are barely covered by Chinese-language tech press. The results are already on Amazon, on the Akutagawa stage, and on Hachette’s preorder lists.

ChatGPT and Claude Aren’t the Right Tools

Most people’s mental model of “writing a novel with AI” stops at one action: open ChatGPT, paste a prompt, let it run.

That approach can’t produce a novel. I’ve covered the reason before: the context window of a large model is a FIFO queue. By round 20, the protagonist’s personality drifts. By round 50, the worldbuilding collapses. Gemini’s 1M-token window sounds huge, but attention tests show that past 50-100k tokens the middle of the context gets ignored — academics call it “lost in the middle.” The window fits, the model just can’t use it.

Authors who actually write long-form have a three-tier answer:

  • Tier 1: Purpose-built front ends for novelists, like Sudowrite and NovelCrafter, which manage context injection for you.
  • Tier 2: Beyond the specialized front ends, use Scrivener or Atticus as the primary writing software, and bring AI in only at specific points — scene expansion, descriptive polish, outline brainstorming.
  • Tier 3: Wire the Claude API into your own workflow (Notion, Obsidian, a custom tool) and treat AI as a callable subtask, not the main writer.

The people using these are almost all Western authors. Let’s walk through them.

Sudowrite Plus Muse Is the Western Mainstream

If you’ve only ever used Claude’s web UI to write fiction, you wouldn’t know: there’s a tool built specifically for novelists that calls Claude, GPT, and Gemini all as backends — and has trained its own dedicated novel-writing model.

It’s called Sudowrite.

In one sentence: it breaks “writing a novel” into dozens of UI components, each backed by a different model call. You’re no longer staring at a blank chat box. You’re inside a Photoshop-style interface — character and worldbuilding panels on the left, the chapter you’re writing in the middle, AI-generated candidate paragraphs on the right.

Sudowrite the company is worth a paragraph on its own:

  • Founded 2020 in San Francisco, bootstrapped, with only one $3M round raised by 2025
  • 16-person team, $1.8M ARR in 2025
  • Investor list reads like a Silicon Valley plus Hollywood crossover: Ev Williams (Twitter, Medium), Sahil Lavingia (Gumroad), Matt Mullenweg (WordPress), John August (Aladdin screenwriter), George Nolfi (director of The Bourne Ultimatum)

In March 2025 Sudowrite launched its in-house model, Muse 1.0. Three months later came Muse 1.5. By 2026, Muse 1.5 is the default. It’s fine-tuned specifically on published novels — bad at coding, bad at trivia, good at exactly one thing: writing fiction.

The company’s changelog claims: in internal blind tests, authors preferred Muse 1.5’s prose to Claude 3.7 Sonnet’s at a 2:1 rate.

Discount that number. The blind test was run by Sudowrite, with no third-party verification and no public methodology. But heavy users on blogs and podcasts keep confirming that Muse really is stronger on sensory detail and prose feel.

An example. Describing “a strange book,” a generic LLM gives you something flat like “the smell of leather.” Muse 1.5 gives you:

“The grind of worn leather against old wood. A thread of metallic blood. The cover carries the loamy smell of soil after a lightning strike.”

That kind of sensory granularity is what Sudowrite paid millions in fine-tuning to get.

Sudowrite’s core features:

  • Story Bible: you write down your worldbuilding, characters, and outline; Sudowrite stores it as structured data and injects the relevant pieces every time you write a paragraph. It’s an always-on novel-setting encyclopedia.
  • Describe: highlight a noun in your text, say “that strange book,” and get 3 candidate descriptions from each of the five senses — sight, smell, touch, hearing, taste.
  • Expand: highlight a skeleton sentence like “She reached out and opened the book,” and it expands into a full paragraph.
  • Brainstorm: drop in a scene or a problem, get 5-10 creative options — no full prose, just sparks.
  • Draft: feed in a chapter outline (200-500 words describing each scene) and generate the full chapter.

What Sudowrite Power Users Actually Do

Feature lists are abstract. Look at three authors who’ve actually shipped books with Sudowrite and how they use it.

Case 1: Joe Vasicek cut his book cycle from 2 years to 14 days

Joe Vasicek is an American indie sci-fi author with 20+ books on KDP, best known for the Captive of the Falconstar series.

He publishes every number from his latest project on his blog:

  • Title: The Soulbond and the Sling
  • Chapters: 18 chapters plus prologue and epilogue, 80 scenes
  • Total word count: 153,254 words
  • Pre-writing phase (outline + worldbuilding): 9 days
  • AI generation phase: 5 days
  • Manual revision afterward: no specific day count given, but “the largest chunk of total time”
  • Sudowrite credits burned: about 770,000

His workflow runs on Claude. He says outright on his blog that Claude’s chapter-level output is “stunning” and makes him “feel more like a first-time reader of the novel than its author.” He uses Muse 1.5 as a parallel generator — same input fed to both Claude and Muse, then he hand-picks the better paragraphs and stitches them together.

His warning to anyone trying AI writing:

“The worst thing you can do is let the AI do all the work. AI doesn’t really create. It analyzes patterns in human language and reproduces them.”

After every AI-generated chapter, he opens a new document and rewrites it by hand. Never copy-paste. The AI draft is a reference he can consult at any time, nothing more.

Case 2: Leanne Leeds writes 12-15 books a year at a 4.5-star Amazon average

Leanne Leeds is an American genre author with 5 series spanning contemporary supernatural, fantasy, and cozy mystery. She has 40+ KDP titles, a 4.5-star Amazon average, and beta readers who don’t notice the AI.

She laid out her full workflow on The Creative Penn podcast:

  • Skeleton draft: she calls it “boring writing.” Just action and dialogue. “He looked at her, she looked at him, he blinked, she kissed him.” No description.
  • Dual-monitor setup: Scrivener on the center screen, Sudowrite plus ChatGPT plus Quillbot on the left.
  • Scene expansion: she selects a skeleton paragraph and uses Sudowrite’s Rewrite to expand it. She gives specific instructions like “show don’t tell” or “describe the look on his face.”
  • Mystery elements: she hands suspect lists and red herrings to ChatGPT.
  • Sentence polish: Quillbot does the final cleanup pass.

The key numbers:

  • AI text as a share of the final draft: about 10% (she measures it with a text-comparison tool, holds steady between 8 and 11%)
  • Time to finish a book: about 3 weeks
  • Every book includes an AI-assisted creation disclosure on the copyright page

Her most-quoted line:

“I use AI to fix the weaknesses in my own writing so that readers get a better experience.”

By 2026 she’s migrated off pure Sudowrite onto a stack of Claude plus Gemini Pro plus NotebookLM. In her latest podcast episode:

“I don’t use AI for plot suggestions — I know where each chapter and scene is going. I use AI to express what I’m trying to say better than I would alone.”

Case 3: Joanna Penn, the AI-Assisted Artisan on the NYT list

Joanna Penn (pen name J.F. Penn) is a British mystery and thriller author whose ARKANE series has hit both the NYT and USA Today bestseller lists. Her 2025 release is Death Valley.

On her own podcast, The Creative Penn, she’s mapped out her positioning: AI-Assisted Artisan Author. The concrete stack:

  • ChatGPT for research and outlining
  • Sudowrite for scene description and expansion
  • Midjourney for covers
  • ElevenLabs to clone her own voice for audiobooks
  • Runway ML for book trailers
  • Every prompt runs over 100 words — “treat it like an employee, give it everything in your head”

In early 2026 she announced a strategic pivot from “high-volume low-price digital books” toward “premium physical books and high-margin products,” with AI handling research and repetitive work while she focuses on creative decisions.

The Second Core Tool: NovelCrafter Plus Codex

If Sudowrite is “an AI writing tool” with AI as the protagonist, NovelCrafter is “writing software plus an AI connector” — you’re the protagonist, AI is what you call.

Company background:

  • Founded 2023, by Leonie Grabel, based in Hamburg, Germany
  • German GmbH, fully bootstrapped, no VC
  • 10-person team (including 2 office dogs)
  • 157,000 registered authors
  • Pricing $4-20/month

The core feature is Codex. In plain English: you build a database of your novel’s worldbuilding, where each entry is a character, a location, a faction, or a prop. Each entry has a name, aliases, a description, custom fields, and relationship links to other entries.

While you write, the system scans the paragraph you’re working on. If the name or an alias of any Codex entry shows up, all of that entry’s information gets dropped into the model’s context automatically. It’s an always-on, keyword-triggered, auto-responding worldbuilding reference manual.

There’s also a cascading relationship feature. If character A is linked to location B, B is linked to faction C, and C is linked to characters D and E, then writing A pulls B, C, D, and E into context automatically.

This mechanism solves what Sudowrite’s Story Bible can’t: cross-book series. A fantasy trilogy shares one Codex across all three books, so protagonist personality, worldbuilding, and faction relationships don’t degrade between volumes.

NovelCrafter is compatible with 300+ models. How? It directly integrates 8 providers (OpenAI, Anthropic, Google, Mistral, Groq, Anyscale, local LM Studio, local Ollama), and the 300+ figure comes via the OpenRouter aggregator. Write one chapter with Claude Opus 4.7, another with locally hosted DeepSeek.

2026 Long-Form Writing Model Benchmarks

inkfluenceai.com recently ran a systematic comparison: 70,000-word novel generation tests scored out of 70.

Key findings:

  • Claude Opus 4.7 is the most stable on cross-chapter character consistency — it “correctly preserves the voice of a minor character who only appears once in chapter 3”
  • GPT-5 is the strongest on plot logic, but loses character detail
  • Gemini 2.5 Pro’s 1M context is theoretically the largest, but it scores lowest on writing style

Then the cost comparison gets brutal:

  • A 70k-word novel via API: Claude Opus 4.7 costs about $32
  • Same book via Claude Sonnet 4.5: about $5.50
  • Quality gap: only 4%. Cost gap: 88%

Community consensus: use Sonnet for routine chapter generation, save Opus for character description review and the key chapters that need polish.

The Scoreboard for AI Writing

Lining up everyone who’s actually written something serious with AI:

  • Rie Kudan: January 2024, Sympathy Tower Tokyo wins Japan’s Akutagawa Prize. About 5% of the text is from ChatGPT. The jury calls it “almost flawless.”
  • Japan’s 13th Hoshi Shinichi Award: roughly a quarter of submissions carried AI signal. The grand prize, Tower of the Genome, was a human-AI collaboration. The award institutionalized rules: no verbatim copying of AI output, mandatory records of the creation process. It’s one of the few global literary prizes that explicitly welcomes AI-assisted submissions.
  • Leanne Leeds: 40+ KDP titles, 4.5-star Amazon average. Beta readers don’t notice the AI.
  • Joe Vasicek: 20+ sci-fi titles on KDP. Book cycle compressed from 6-24 months to 1-2 months.
  • Joanna Penn: ARKANE series on the NYT and USA Today bestseller lists.
  • Elisa Shupe: American military veteran, AI Machinations. In April 2024 the U.S. Copyright Office granted limited copyright to an AI-collaborative work for the first time — a landmark case in copyright law.

Readers Can’t Tell Anymore

This is one story, and it needs telling from the start.

February 2025: an American author named Mia Ballard self-published a horror novel on Amazon KDP called Shy Girl. The cover was a girl’s silhouette; the themes were trauma, isolation, childhood shadow.

The book caught fire on BookTok (TikTok’s reading community). Book reviewers filmed unboxing videos of the physical edition, gave it 5 stars, recommended it enthusiastically. On Goodreads it racked up over 4,900 ratings, averaging 3.52 stars.

The buzz reached Hachette, one of the Big Five publishers — meaning it sits next to Penguin Random House, HarperCollins, Simon & Schuster, and Macmillan at the very top of global publishing. Hachette’s UK imprint Wildfire is the one that published Fifty Shades of Grey.

In March 2025, two Hachette imprints — Orbit Books in the US and Wildfire in the UK — jointly acquired the rights to Shy Girl. A self-published horror novel had walked all the way from KDP to a Big Five preorder catalog.

In November 2025 the UK edition shipped about 1,800 copies.

Up to this point it’s a textbook BookTok success story: viral indie hit gets picked up by a major publisher.

The turn came in January 2026.

A YouTube book reviewer named Frankie (channel: Frankie’s Shelf) dropped a 2-hour 40-minute analysis video titled “i’m pretty sure this book is ai slop.”

Her specific evidence:

  • Repeated use of the “three in a row” structure when describing the position of door locks
  • An unusually high frequency of the word “sharp”
  • Layout choices that were “baffling,” including jarring paragraph breaks
  • Inconsistent narrative voice across passages — like different people had written them

The video went viral on BookTok and X, hitting 1.5M+ views.

What came next was sharper. Thad McIlroy, a publishing industry analyst, submitted the Shy Girl ebook to Pangram, an AI detection company. Pangram is one of the more academically cited AI detection tools today.

Pangram’s verdict: 78.4% of the text was AI-generated.

McIlroy passed the report to New York Times reporter Alexandra Alter. On March 19, 2026, the NYT ran the investigation.

Within a day of being notified, Hachette canceled the U.S. publication and halted UK sales. The statement was carefully worded: after “an extensive investigation,” they had decided to cancel publication. No direct accusation of the author.

Mia Ballard’s response was unexpected. She denied using AI to write the book — claiming that “a freelance editor she’d hired had used AI tools without her knowledge.” Citing ongoing litigation, she declined to comment further.

This is publishing’s defining event of 2026, because it exposed a loophole no one had thought about: existing publishing contracts only require AI disclosure at the author level, not from editors, proofreaders, literary agents, or anyone else in the production chain. A book might not be written by the author using AI; it might be edited by the editor using AI.

And the 4,900 BookTok readers who gave it 5 stars never noticed. Hachette’s acquisition editors never noticed. Not until one YouTube reviewer spent 2 hours and 40 minutes dissecting the text, and one AI detection company put a 78.4% number on it, did the story break.

Where the Big Five Stand on AI

None of the Big Five publishers has issued a company-wide policy banning AI use by authors. That policy vacuum is one root cause of the Shy Girl incident.

Amazon KDP’s policy is delicate:

  • It distinguishes “AI-generated” (must be disclosed) from “AI-assisted” (no disclosure needed)
  • Grammar checks and rewriting suggestions count as assistance
  • Disclosures are not shown to readers on the product page, only kept in Amazon’s internal records
  • Enforcement ramped up in 2025-2026, but no names of banned AI-author accounts have been made public — enforcement is mostly quiet delistings

U.S. Copyright Office, March 2026: the Supreme Court declined to hear the case on whether AI alone can create a copyright-protected work, upholding the lower court ruling — AI cannot be a copyright holder.

The Authors Guild launched a Human Authored certification program, $10 per book for non-members. Their 2025 survey:

  • 90% of authors believe they should be compensated if their work is used for AI training
  • 96% demand it requires consent

Where the Chinese-Language Writing World Stands

Start with the counterexamples. The Chinese-speaking world isn’t a total void on AI writing, but everything sits at one of two extremes: academic stunts or web-novel assembly lines.

Academic experiments:

  • Shen Yang’s team at Tsinghua University’s School of Journalism used AI to generate a ~40,000-word sci-fi novella, Land of Machine Memory, in 3 hours (66 prompts, Kafka-style). It won second prize at the 5th Jiangsu Sci-Fi & Popular Science Award, submitted anonymously. After the team came forward, Liu Cixin said it was “honestly better than what I’d write” and felt “a huge sense of loss.”
  • Wang Feng’s team at East China Normal University used AI to write a ~1.1M-word novel, Apostle of Fate, over 6 weeks.

Neither was a commercial release. Both were academic demos.

Web-novel assembly lines:

  • Tomato Novel (Fanqie) hit 5,606 new-book debuts in a single day in March 2024, up 13x from 400 in 2023 — pure AI-driven mass production
  • April 2025: Tomato tightened its signing process
  • March 2025: Qidian banned all AI-generated content
  • Jinjiang allows proofreading and polishing, but bans AI-driven plotting
  • No Chinese platform allows AI-labeled works to be displayed or distributed on their own

Chinese AI writing tools are almost entirely web-novel focused: Waqu Pinwen and Wawa Writing (Hangzhou Yinli Zhihang Tech) pitch long-form web-novel memory, “removing the AI smell,” and outline generation. Yuewen’s Author Assistant integrates DeepSeek-R1. There is no Chinese equivalent of Sudowrite — no professional tool aimed at serious long-form fiction.

Where Chinese literary heavyweights stand:

  • Mo Yan: AI prose is impressive but “lacks real thought or creativity”
  • Liu Cixin: read Land of Machine Memory, said “honestly better than what I’d write,” felt a sense of loss — doesn’t use AI himself
  • Han Song: AI can be a collaborator that sparks ideas, but the unique human experience and subconscious insight aren’t things AI has
  • Mao Dun Literature Prize and Lu Xun Literary Prize: rules don’t explicitly mention AI but emphasize originality
  • Science Fiction World magazine: explicitly stopped accepting AI-generated fiction in 2023

The Chinese writing world is a blank space on “an indie author using AI assistance to write a quality long-form novel.” Not for one single reason:

  • Misaligned commercial incentives: industrialized web-novel mass production is far more profitable than serious AI-assisted literary work
  • Model quality: Chinese-language LLMs lag Claude and GPT-5 on long-form coherence and emotional depth
  • Cultural taboo: publicly admitting AI use in China can mean delisting, copyright disputes, social-reputation pressure
  • Tool gap: every domestic tool targets web-novel volume; nothing supports serious long-form workflow
  • Distribution gap: China has no Amazon KDP — no direct self-publish channel to readers

Stack those five and Chinese-language AI novel writing only exists at two poles: the bottom rung of the web-novel assembly line and the academic lab. The middle — indie authors using AI to produce quality long-form work — is empty.

What the Truth About AI Writing Looks Like

Put all the data together and here’s where AI writing actually stands:

1. AI doesn’t turn bad writers into good ones. Joe Vasicek, Leanne Leeds, Joanna Penn, Rie Kudan — the people producing real long-form work were already writers. AI made them 5-10x faster. It didn’t raise their ceiling.

2. AI massively lifts the productivity of good writers. A 6-month-to-2-year book cycle compressed to 1-2 months. A KDP genre author shipping 12-15 books a year, 4.5 stars average. This is happening, not predicted.

3. AI writing tools keep getting stronger. Sudowrite, NovelCrafter, Claude API plus Scrivener are putting publishable work on shelves every day.

4. Quality still isn’t there. Claude Opus 4.7 scored 63.2 out of 70 in benchmarks — the highest combination today — but no case yet of “AI as primary writer produces a lasting work of literature.” Rie Kudan’s Akutagawa novel was only 5% AI.

5. Readers already can’t tell. Shy Girl had 4,900+ Goodreads ratings at 3.52 stars on average. Hachette’s acquisition editors didn’t catch it. Not until a YouTube reviewer spent 2 hours and 40 minutes on the text plus an AI detection company posted a 78.4% number did anyone know. AI fiction is already clearing the average reader’s taste filter.

6. The rules of the game are still being written. Hachette pulling Shy Girl exposed the Big Five’s policy vacuum. Amazon KDP’s distinction between pure AI generation and AI assistance leaves an enormous gray zone. The U.S. Copyright Office’s stance shifts monthly.

7. The Chinese-language scene lags. The cause is a four-way absence: culture, commerce, regulation, and tool ecosystem.

Problems keep getting solved. AI-written novels keep going on sale. Readers keep handing out 5 stars and can’t tell whether or not it’s AI.

I’m starting to wait for the first AI novel that lasts.