A book is the longest single piece of prose most authors will ever ship. It runs 50,000 to 90,000 words across 8 to 20 chapters, often written over many months. If chapter three reads natural and chapter four reads like a ChatGPT default, the reader feels the gap and stops reading. Beta readers leave the manuscript at the same break. Literary agents reject the same flatness in the first ten pages. Big Five imprints (HarperCollins, Penguin Random House, Simon and Schuster, Hachette, Macmillan) now run AI-screening on submitted manuscripts, and Amazon KDP requires AI disclosure at upload. TextSight scans each chapter, flags the AI tells specific to long-form fiction and non-fiction, and rewrites the flagged lines back into the voice readers signed on for. Built for literary fiction, genre fiction (thriller, romance, sci-fi, fantasy), memoir, business and self-help non-fiction, and academic non-fiction.
Reader DNF rates, agent pass rates, and acquisition-editor screening are three signals on the same underlying thing. They all ask whether chapter four sounds like the same writer who wrote chapter one. ChatGPT defaults erode that signal slowly, which is why authors notice it months after a draft has set. The honest framing is that rewriting is not about hiding ChatGPT involvement. It is about keeping the voice readers, agents, and editors signed on for.
By chapter three, a reader has internalised your sentence rhythm, your habitual hedges, the way you handle dialogue tags, the kinds of small specifics you tend to drop in. When chapter four lands in a different register, they cannot always name what changed, but the trust calculus shifts. The cheapest way to keep that calculus stable is to keep the voice stable, and the cheapest way to keep the voice stable when half the chapters were drafted in ChatGPT is to rewrite before the manuscript goes to anyone, beta reader or agent.
Goodreads, Amazon, and beta readers do not run AI detectors. They just stop reading. The review they leave uses words like flat, textbook-y, recycled, or AI-sounding without naming a cause. Self-published authors on KDP report that DNF rates and one-star reviews climbed through 2024 and 2025 as readers got faster at recognising AI prose by feel. A chapter does not need to be detectably AI to lose the reader, and the comments on the review page do not specify a fix.
Authors who started a manuscript in March and finished in October find their natural voice has drifted in eight months even without AI involvement. Add ChatGPT to half the chapters and the drift becomes a chasm. The reader notices the chasm at exactly the wrong moment, usually around chapter four. That is the structural inflection point where genre fiction earns the next forty pages and non-fiction earns the next argument. Voice consistency across the manuscript is the discipline that keeps the reader past that inflection point.
These are the patterns the detector flags most often in ChatGPT-drafted book chapters, and they are the ones experienced readers and beta readers learn to recognise across two or three chapters. None is wrong in isolation. The pattern across a chapter is what reads AI.
ChatGPT defaults to a small set of memoir and fiction openers built around childhood reflection and origin-story framing. "Ever since I was a young girl," "From the moment I could remember," "Growing up in a small town." None is bad writing in isolation, but they signal generic-narrative training data immediately. The AI rewriter flags origin-story openers and suggests dropping the reader inside the scene the chapter actually contains instead.
Real fiction and memoir vary paragraph length sharply. A two-word paragraph lands a beat. A 200-word paragraph carries an unbroken thought. ChatGPT defaults to paragraphs of roughly 80 to 120 words, all of similar internal rhythm, all closing on a similar cadence. The pattern is invisible at the sentence level and obvious at the page level. The AI rewriter flags rhythm uniformity and suggests where to break a long paragraph into a single-line beat or fuse two short ones.
ChatGPT closes nearly every chapter on a summarising paragraph that restates the chapter's themes in slightly different language. Trade fiction and modern non-fiction rarely do this. They close on the last specific scene, claim, or unresolved tension and trust the reader to carry it. The AI rewriter flags summary closers and suggests cutting them entirely, letting the chapter end on its last substantive line.
"She felt a wave of relief wash over her." "His heart pounded in his chest." "A knot formed in her stomach." ChatGPT defaults to a stock vocabulary of physical sensations for emotional states. Real fiction writers vary the embodiment, often skipping it entirely, often anchoring it in a specific gesture instead. Romance, thriller, and literary fiction readers spot these stock beats faster than any other tell. The AI rewriter flags them and suggests replacing with a specific gesture or a beat of dialogue.
ChatGPT clusters dialogue tags ("she said," "he replied," "she whispered," "he murmured") and reaches for an adverb modifier on roughly every third tag. Modern fiction trims tags aggressively, uses "said" most of the time, and lets context carry the speaker. The AI rewriter flags dialogue-tag density and adverb-modifier patterns. This is the most fragile tell because over-correcting it can confuse who is speaking, so we recommend Light mode on dialogue-heavy chapters.
ChatGPT defaults to a three-act mini-arc inside every chapter (setup, complication, resolution). By chapter five the reader feels the rhythm before they read the prose. Real chapters vary structurally. Some land on a finding, some on an unresolved question, some on an anecdote, some on a single image. The AI rewriter flags the three-act pattern and suggests breaking it on at least three chapters in the manuscript.
In non-fiction, ChatGPT subheadings follow the same templates across chapters: "The X You Need to Know," "Three Ways to Y," "Why Z Matters." Real trade non-fiction subheadings are specific and sometimes weird. They name the thing the section is actually about. The AI rewriter flags template subheadings and suggests rewriting each one to name the concrete subject, even if the result is less clickable.
Big Five imprints and most established indie presses now run AI-screening on submitted and acquired manuscripts. Amazon KDP requires AI disclosure at self-publishing upload. The honest framing for authors is that authenticity is a craft step for reader experience, not a compliance workaround. Disclose what you used, then rewrite for the reader.
HarperCollins, Penguin Random House, Simon and Schuster, Hachette, and Macmillan run internal AI-screening passes as part of editorial workflow. Some screen at the submission stage, others at the contract stage, several at both. The screening tools are not always disclosed by name, and the thresholds vary by imprint. What is consistent is that a manuscript reading detectably AI-flavoured raises a flag the acquisition editor has to defend internally before the book can move. A clean voice score removes that friction entirely.
Indie literary presses and university presses screen less systematically but care more about voice. A manuscript that reads polished-but-voiceless is the most cited pass reason from this segment, with or without AI screening. The fix is the same as for the Big Five: rewrite each chapter into your established voice, particularly the connective prose between cited material or interviews.
Amazon requires self-published authors to disclose AI-generated content during KDP upload as part of their content policy. The form asks two questions. Is the text AI-generated (created by AI with limited human editing). Or is it AI-assisted (you wrote the prose yourself with AI helping you brainstorm or copy-edit). Rewriting an AI-drafted chapter does not change the underlying answer. If ChatGPT drafted the prose, the honest answer is AI-generated even after an authenticity pass. Disclose honestly, then rewrite for reader experience.
Indie press and trade publisher contracts increasingly include AI disclosure clauses in the author warranty section. Read the warranty before signing. If your contract requires original human work and you used ChatGPT for substantive prose, you have a contractual problem the AI rewriter cannot solve. Use AI for brainstorming and outlining when the contract requires original prose, and disclose openly when the contract permits AI assistance with disclosure.
For book chapters, the mode depends on what the chapter contains. Dialogue carries character voice and is the most fragile to aggressive rewrites. Narration tolerates more rewriting. Maximum is risky on any short paragraph because aggressive rewrites can shift meaning and erase the specific phrasings that make a chapter yours.
Light makes mild edits and preserves sentence structure, dialogue tags, character voice, and personal anchors. Score gains per pass are smaller, but the output still reads like the same writer wrote it. Use Light on dialogue-heavy chapters in any genre, on romance generally, on first-person memoir scenes, and on any chapter where the prose was already substantially yours and the ChatGPT contribution was limited to a section or two.
Balanced runs moderate rewrites and is the right choice for most narration, exposition, business non-fiction, and academic chapters. It shifts cadence and vocabulary enough to clear the seven tells without flattening voice or erasing the specific phrases that make the chapter recognisable. For a 5,000-word ChatGPT-assisted chapter, a Balanced pass on the red sentences typically moves the Authenticity Score from the 10 to 30 range into the 70 to 85 range.
Maximum runs the most aggressive rewrite and produces the biggest score gain per pass. For book content the caveat is real: aggressive rewrites can shift the meaning of short paragraphs and erase the personal anchors that distinguish your voice across a manuscript. Reserve Maximum for isolated red sentences after a Balanced pass and always read the rewritten line in the context of the surrounding paragraph, not just the sentence, before accepting the rewrite.
The default we recommend for book chapters: Balanced on narration, Light on dialogue, Maximum only on stubborn red lines that did not clear after a Balanced and a manual edit. Treat the manuscript as twelve separate authenticity jobs, not one.
All three modes available on every paid plan. Pro is $19.99 monthly or $14.99 effective on yearly. Business adds API access and team seats. Full details on the pricing page.
Billed $89.88/year — Save $30
Billed $179.88/year — Save $60
Billed $359.88/year — Save $120
Yearly billing saves 25%. View full pricing →
A 5,000-word chapter is roughly 30,000 characters. Pro at $19.99 monthly or $14.99 effective on yearly gives you 50,000 AI rewriter words per month, enough for one to two chapters per session. A 12-chapter manuscript takes two to four weeks at a sustainable pace. Here is the per-chapter loop authors converge on.
Paste the chapter into TextSight before you change anything. The scan returns an Authenticity Score (typically 10 to 30 for unedited ChatGPT drafts) and a sentence-level colour map. Save the baseline. Any chapter below 60 usually needs structural rework before rewriting is worth doing. Between 60 and 75 is the rewriting zone. Above 75 you are polishing.
Before running the auto-AI rewriter, do a pass for the seven book-chapter tells. Rewrite any origin-story or "ever since" opener. Break up uniform paragraph rhythm. Cut summary closers. Replace generic emotional beats with specific gestures. Trim dialogue-tag clusters and adverb modifiers. Break the three-act pattern on at least one section. Rewrite template subheadings. These edits alone usually lift the score by 15 to 25 points.
Balanced on narration and exposition. Light on dialogue. Light on first-person memoir scenes. Maximum only on stubborn red lines that did not clear after a Balanced and a manual edit. Apply mode per passage, not per chapter. A chapter with dialogue and narration may need both Light and Balanced runs on different paragraphs.
Literary fiction. Voice is the product. Balanced on narration, Light on dialogue, Maximum only on stubborn red sentences after a Balanced pass. Plan on heavier authenticity than any other category. Genre fiction (thriller, romance, sci-fi, fantasy). Readers forgive a lot for pace, but flat dialogue and generic emotional beats break the spell faster than any other tell. Light on dialogue, Balanced on action and exposition. Romance dialogue is the most fragile, Light only. Memoir and literary non-fiction. Memoir is voice plus specificity. The origin-story opener is the most common tell. Balanced on narrative sections, and replace any generic emotional beat with a concrete sensory detail from your actual memory. Business non-fiction and self-help. This shelf is saturated with ChatGPT drafts. Balanced on connective prose, leave case studies untouched, and sharpen opinion sentences from "it can be argued that" into committed claims. Academic non-fiction. Vary chapter intros, break the three-act pattern on at least three chapters, preserve quotes and citations exactly. Balanced on connective passages, Standard preservation on cited material.
Open the chapter you wrote first without AI assistance. Find three phrases you used naturally, the connectives or hedges or quirks you reach for under pressure. Plant at least one of each in the rewritten chapter. This is what makes the voice consistent across 200 pages rather than just locally human. Without this step, every chapter ends up locally clean but cumulatively neutral.
In non-fiction, an experience anchor is a real number, a real client name with permission, a real failure, an opinion the model would not commit to. In fiction, the anchor is a specific sensory detail, a piece of overheard dialogue, a small object that does not need explanation. ChatGPT cannot fake these. The prose pattern around them reads obviously human.
Re-scan the full chapter and confirm the Authenticity Score sits at 70 or higher (80+ for any chapter going to an agent's first-ten-pages sample). Send three chapters at a time to your beta readers. When their feedback comes back, map "felt generic" or "lost interest" comments to specific chapters in your scan history and rework only those, not the whole manuscript.
Books are a relationship business between author and reader, and the publishing industry takes AI provenance increasingly seriously. We want to be explicit about what this tool is for and what it is not.
Manuscripts you researched, outlined, and shaped yourself, with ChatGPT used as a drafting assistant for some chapters or passages. The structure is yours, the arguments are yours, the experience anchors are yours. The AI rewriter helps you catch sentences where the assistant register leaked into the prose so the final version reads in your established voice. This is closer to a careful proofread than to laundering, and it is the only ethical use of an editorial AI rewriter on book-length work.
Generating a complete novel or non-fiction book from ChatGPT prompts and running it through any AI rewriter to mass-publish on KDP or query agents. The signal readers and acquisition editors eventually pick up on is not just prose patterns but the absence of original thought across chapters. Rewriting does not produce original thought. It also does not retroactively change the answer to KDP's disclosure question. The right framing is that rewriting preserves the voice of original work, not that it manufactures the appearance of original work.
Read your author warranty clause before signing. Several Big Five contracts now require disclosure of substantive AI use, and a few prohibit it for the prose itself. If your contract requires original human work and you used ChatGPT for substantive prose, you have a contractual problem the AI rewriter cannot solve. Use AI for brainstorming and outlining when the contract requires original prose, and disclose openly when the contract permits AI assistance with disclosure. The AI rewriter is part of the editorial layer that keeps your voice consistent — not a substitute for the rest of that layer.
The detector page focused on acquisition editors, slush-pile screening, and publishing-house workflow.
Open the detector →The sibling page for newsletter operators: voice consistency, paid-subscriber trust, open-rate aware rewrites.
Open newsletter AI rewriter →The flagship AI rewriter page covering all source content. Three modes, closed-loop calibration, no signup.
Open AI rewriter →How the score is computed and what threshold to aim for before you query an agent or upload to KDP.
Read the guide →Free to try, no card. Sentence-level highlights, three modes, voice-led rewrite, KDP-disclosure safe.