Why AI-Written Books Sound Generic — and How to Keep Your Voice

You asked an AI model to draft a chapter on the thing you teach every week. What came back was tidy, well organised and grammatically perfect. It was also strangely empty. Your strongest opinion had been softened into “it depends”. The client story was one you never told. Two of the points had already appeared, almost word for word, in the previous chapter.

Nothing in it was exactly wrong. The problem is that a competitor could have put their name on it, and a client who knows you would not recognise it as yours.

That complaint is common among coaches and consultants who have tried drafting a book with ChatGPT or Claude. But “generic” is not a mystery. It has a small number of causes, and each leaves a recognisable mark on the page.

Generic is a symptom, not a style choice

Stop treating generic prose as a tone problem you can prompt away with “make it more human”. A language model writes the most plausible continuation of what it has been given. Given little, that continuation is the most common version of your topic: the average of the internet.

The ghostwriter and Georgetown writing lecturer Michael Long put it neatly at the 2025 Gathering of the Ghosts: “Everything you leave out will be filled in with blandness” (ASJA). In the same report, Katia Walsh, co-author of Winning with AI, calls AI prose “incredibly bland”, full of clichés and corporate jargon, and the report notes that the ghostwriters who use AI do not use it to write.

For a coach, this matters more than for most authors. Your book is supposed to show a prospective client how you think. A book that sounds like everyone else’s proves only that you can publish.

Six reasons why AI writing sounds generic in a book

Each cause below has a symptom you can see on the page, a quick test and a fix. Most bland chapters suffer from two or three at once.

1. No source material, so the model writes the typical version

Symptom: The chapter covers the obvious sub-points of the topic in the obvious order. Nothing in it surprises you, including the parts you disagree with.

Test: Highlight every sentence that only you could have written, because it rests on your method, your client work or your experience. If little is highlighted, the model had nothing of yours to draw on.

Fix: Give the model your actual material for that chapter: talk transcripts, workshop slides, client notes (anonymised), old articles, voice memos. If you have years of material scattered across folders, sort it before you draft; our guide to turning years of coaching experience into a structured book covers that step.

2. No voice evidence, only adjectives

Symptom: You asked for “conversational, direct and warm” and got prose that performs those adjectives: rhetorical questions, “let’s be honest”, a joke that is not one you would make.

Test: Read a paragraph aloud next to a paragraph you wrote yourself. Compare sentence length, how sentences open, how often you use “I”, and whether you make claims or soften them.

Fix: Replace adjectives with samples and observable features. Anthropic’s own prompting guide calls examples “one of the most reliable ways to steer Claude’s output format, tone, and structure” (Anthropic). Describe what is measurable: “short declarative sentences, rarely more than two clauses; first person; no rhetorical questions; states a position, then the exception”. Our article on training AI on your writing style for a whole book goes through this in detail.

3. The model’s default register pulls towards safe and balanced

Symptom: Every strong claim is hedged. Paragraphs end with a balanced summary (“Ultimately, the right approach depends on your situation”). Vocabulary is polished and slightly corporate.

Test: Find the three sentences in the chapter a reader could argue with. If you cannot find three, the chapter has no position.

Fix: State your positions explicitly before drafting and tell the model not to soften them. There is research behind this pull towards sameness. In a study of argumentative essays, writing with InstructGPT, a model tuned on human feedback, significantly reduced content diversity and made different authors’ essays more alike, while the base GPT-3 model did not; the reduction came mostly from the model’s contributions, not the text people wrote themselves (Padmakumar & He, ICLR 2024). Kirk and colleagues report that reinforcement learning from human feedback “significantly reduces output diversity” compared with supervised fine-tuning (Kirk et al.). And in a short-story experiment, stories written with AI-generated ideas were rated better written but were “more similar to each other than stories by humans alone” (Doshi & Hauser).

The style is recognisable at scale, too. An analysis of more than 15 million PubMed abstracts found an abrupt increase in the frequency of certain style words after LLMs arrived (Kobak et al.). Your readers may not be running word counts, but they have read a great deal of that register by now.

4. One chapter at a time, with no memory of the book

Symptom: Chapter 6 makes the point Chapter 3 already made. Your framework has three stages in one chapter and four in another. Each chapter opens with a scene-setting paragraph and closes with a recap.

Test: List the main claim of each chapter in one sentence. If two sentences are near-duplicates, or a framework name changes, you have drift.

Fix: Draft from a whole-book plan: the book’s central argument, a one-line job for each chapter, and a glossary of your terms and framework names. Paste the relevant parts into every drafting session. Even within a single long conversation, models do not use everything equally well; one widely cited 2023 study found performance dropped when the relevant information sat in the middle of a long input (Liu et al.). Do not rely on the model remembering chapter 2 when it writes chapter 9.

5. Invented anecdotes replacing your real ones

Symptom: “Consider Sarah, a marketing director who struggled with…” The story is neat, resolves in four sentences, and never happened.

Test: For every example, client result and statistic, write down where it came from. Anything you cannot trace is invented, however plausible it sounds.

Fix: Tell the model never to create examples, client results, quotes or figures, and to mark a placeholder (“[client story needed: a manager who over-delegated]”) instead. Then fill the placeholders from your own case notes. An invented case is not just bland; if a former client thinks it is about them, it is a trust problem. The main-site guide on editing an AI-assisted book includes a fact-checking pass for this.

6. You outsourced the thinking, not just the drafting

Symptom: The chapter’s argument is reasonable but not one you would actually make in a client session. It reads like a summary of the field.

Test: Ask yourself whether you would say this chapter’s main point, unprompted, to a client who was paying you. If you hesitate, the position is the model’s, not yours.

Fix: Decide the argument yourself. Write a few rough sentences per chapter on what you believe, what you disagree with in common advice, and which client situation proves it. AI can then help shape and draft around that thinking. It cannot supply it.

The diagnosis at a glance

Cause	What you see on the page	Quick test	Fix
No source material	Obvious points in the obvious order	Highlight sentences only you could write	Draft from your transcripts, notes and slides
No voice evidence	Performed tone: forced questions, jokes	Read aloud next to your own paragraph	Samples plus measurable style features
Default safe register	Hedges, balanced recaps, polished jargon	Find three arguable sentences	State positions; forbid softening
No whole-book memory	Repeated points, drifting framework names	One-sentence claim per chapter	Book plan, chapter jobs, term glossary
Invented anecdotes	Neat stories that never happened	Trace every example to a source	Placeholders, then your real cases
Outsourced thinking	A summary of the field	“Would I say this to a client?”	Write your positions first

Before and after: one paragraph on delegation

The difference is easiest to see side by side. Both paragraphs below are illustrations we wrote for this article. The “after” version voices a hypothetical leadership coach; the coach and the story are invented to show the technique, not a real client.

Before (typical AI-style draft):

Delegation is one of the most important skills a new manager can develop. By empowering team members to take ownership of tasks, managers can free up time to focus on strategic priorities. However, delegation can be challenging, and many new managers struggle to let go of control. The key is to start small, communicate expectations clearly and provide support along the way. Ultimately, effective delegation benefits both the manager and the team.

After (grounded in a hypothetical coach’s position and story):

Most new managers I coach can hand over a task perfectly well. What they cannot hand over is the decision attached to it. Picture an operations lead who gives her analyst the weekly supplier report, then rewrites the recommendation section every Friday for two months. By the third week the analyst has stopped writing recommendations at all. She has not delegated anything; she has hired a typist. So my rule for new managers is blunt: if you are going to rewrite the conclusion, keep the work. If you give it away, give away the right to be wrong with it.

What changed is not vocabulary. The second paragraph takes a disputable position (the problem is delegating tasks, not decisions). It uses a situation with details you can picture (Friday rewrites, week three). It ends on a rule in the author’s phrasing instead of a balanced recap. None of that comes from a better prompt. It comes from the author.

A ten-minute diagnostic for any chapter

Run this on one chapter before you draft the rest. If it flags several items, fix the inputs rather than editing the prose.

What to write yourself and what AI can carry

A short version, because the split deserves its own article. Keep for yourself: the book’s central argument, your positions and disagreements, your framework and its names, the choice of client stories, and the final read of every chapter. AI can reasonably help with organising raw material, proposing structure for a chapter you have already thought through, turning transcripts into first-draft prose, spotting repetition, and tightening sentences.

If you would rather not use AI for drafting at all, that is a legitimate choice; our guide to writing a business book without a ghostwriter covers the other routes, and how to write a thought leadership book covers developing the positions in the first place.

Common mistakes when trying to make AI writing sound human

• Piling adjectives into the prompt. “Witty, authoritative, warm, punchy” gives the model a costume to wear, not a voice to match.
• Reaching for a humaniser. These tools are designed around AI detectors. Your aim is to sound like yourself to a reader who knows you, which is a different job.
• Asking for “more personal” prose. Without real stories to draw on, the model invents personal-sounding ones. That is how the Sarah anecdote gets in.
• Line-editing a chapter with nothing in it. If the diagnostic shows missing material, polishing the sentences wastes your time.
• Accepting placeholders as content. “We’ll swap in a real example later” has a way of reaching the printer.
• Never reading the book as a whole. Chapter-by-chapter review cannot catch repetition and drift across ten chapters.

Where guided software fits, and where it does not

General chat tools can work if you supply the material, samples and plan yourself. Guided software is one alternative. WritebookAI, for example, analyses your documents, notes and transcripts to show what they cover and what is missing, then runs a verification interview on gaps and conflicts. Chapters are drafted from your material and confirmed answers with claim checks; when something is missing, such as a client result or an unsourced statistic, the chapter leaves it out and asks you rather than inventing it. Its Writing DNA profile is built from your own samples, flags where the evidence is thin, and is tested on a fresh topic before you approve it. None of that removes the need for your review: you still check claims and approve every chapter, and it is not built for fiction or for people who want the book written for them. More at writebookai.com.

What to do with the chapter on your desk

Run the checklist on the bland chapter you already have and note which cause shows up most. If it is missing material or positions, stop editing, write down what you believe and which client situation proves it, then redraft. If it is voice, gather real samples and test them on a topic they do not cover. If it is drift, write the one-page book plan before you touch chapter two again.

Generic output is what an AI model produces when the author is absent from the inputs. Put yourself back in before the draft, and you will spend far less time trying to edit yourself back in afterwards.