The Writing System 9 min read

How to Train AI to Write Like You (and Test Whether It Does)

Training AI on your voice usually means good samples and a written voice profile, not building a model. How to choose samples, build the profile, prompt with it and run a blind test.

For most people, training AI to write like you does not mean building a model. It means giving an assistant good samples of your writing and a written description of your voice, then testing the output against your own work until a reader who knows you cannot easily tell them apart. The test is the step most people skip, and it is the only way to know whether the training worked.

This guide covers training and testing a voice in general. For the workflow of drafting a whole book this way, see writing a book in your own voice with AI.

What “training” means in practice

Large language models already know how to write. Whether you use ChatGPT, Claude or another assistant, “training” it on your voice means steering it toward your patterns, not teaching it to write from scratch. There are three levels, and most people only need the first two.

Level What you do Good for Limits
1. Instructions plus samples Paste samples and instructions into a chat, or save them as standing instructions or project files if your assistant offers that Posts, emails and other short pieces Samples get buried in long conversations, and the assistant may copy surface phrases
2. A written voice profile Turn your samples into observable traits, with evidence and anti-examples, and use the profile every time Longer work and consistent results across sessions Takes an hour or two to build and still needs testing
3. Fine-tuning Adjust a model on many examples of your writing through a provider’s developer tools Organizations producing large volumes in one fixed style Needs a lot of clean text, technical setup and budget; rarely worth it for one person’s voice

Most assistants offer some version of level 1, and some providers offer fine-tuning to developers. The details differ by product and change often, so check what yours supports today. The rest of this guide focuses on level 2, because it works with almost any tool and it forces you to say what your voice actually is.

Choose samples that are really you

The samples set the ceiling. If they are not your real voice, nothing downstream fixes it. Good samples are:

  • Written without AI. Otherwise you are training the assistant to imitate itself.
  • Written for real readers. Client emails, posts, articles, proposals and talk transcripts. Not school essays, legal text or reports written to a template.
  • From several contexts. Something explanatory, something persuasive, a story, and a transcript of you speaking if you have one. Speaking often shows your natural rhythm better than polished writing does.
  • Mostly yours alone. Pieces heavily edited by a colleague, a communications team or a ghostwriter carry their voice too.
  • Recent and in the register you want. If you want the AI to write a business book, a decade-old academic paper is the wrong sample.

As a guide rather than a rule, aim for several thousand words in total. Fewer, and the AI fills gaps with generic choices. More helps only if it adds variety.

If most of your writing lives on social platforms, turning LinkedIn posts or a newsletter into a book covers getting it out and which pieces make good samples.

Build a voice profile from evidence

A voice profile describes what your writing does, points to evidence, and says what to avoid. Adjectives such as “warm, authoritative and witty” give the AI nothing specific to do, so it falls back on a stereotype of those words. Observable traits work better.

The table below is an illustration for a fictional operations consultant, not a template to copy.

Dimension Evidence from my writing Use this Avoid this
Sentence length Mostly short to medium; long sentences only for lists of conditions Vary length; most sentences under about 20 words Runs of long sentences with stacked clauses
Openings Starts with the claim or a specific scene “Most handoffs fail in the first week.” Openers that restate the question or start with “In today’s world”
Vocabulary Plain words; industry terms defined once “margin”, “handoff”, “owner” “synergy”, “holistic”, “robust”
Explanation Rule, then reason, then a client example “Do X. The reason is Y. One client…” Abstract principles with no example
Stance Takes a clear position and names the exception “Cancel the meeting. The exception is…” Hedging every claim, or “it depends” without saying on what
Humor Dry, rare, usually at his own expense One aside per piece at most Puns and exclamation marks
Paragraphs Three to five sentences Paragraphs that each make one point One-line paragraphs for emphasis
Endings Ends on the practical next step “Start with the one handoff that costs most.” Inspirational summary lines

Add two or three short anti-examples: passages written in a voice that is close to yours but wrong, labeled “not like this”. They show the AI where the boundary is better than any adjective.

To build the profile, read your samples with a highlighter and note what repeats: how you open, how long your sentences run, the words you use and the words you never use, how you handle disagreement. If you find it hard to see your own patterns, ask a colleague to read three samples and describe how you write, and write down their observations as traits.

Prompt with the profile

A prompt that uses the profile well has five parts:

  1. The voice profile, in full.
  2. Two or three samples, labeled as reference for voice only, not for content.
  3. The task: who the reader is, what the piece is for, the points to cover, and the length.
  4. The content itself: your notes, bullet points or a transcript. AI writes in your voice best when the ideas are yours. Voice without your substance becomes an imitation of your manner wrapped around generic advice.
  5. Constraints: do not reuse sentences from the samples, do not add facts, stories or numbers that are not in the notes, and mark anything you had to guess.

Using the voice profile and samples above as a guide to style only, write a 400-word section for [reader] that explains [point], based only on the notes below. Do not copy sentences from the samples. Do not add examples, statistics or quotes that are not in the notes. Put [CHECK] next to anything you were unsure of.

Review the first draft against the profile line by line, then add what you corrected to the profile. Each round makes the next draft closer.

Run a blind test

Your own judgment is not enough, because you know which paragraphs the AI wrote. Run a blind test instead.

  1. Choose a topic you have not written about, so the AI cannot borrow from your samples.
  2. Write three paragraphs yourself, or use three recent paragraphs that are not among your samples.
  3. Give the AI the same notes and have it write three paragraphs.
  4. Remove formatting differences, shuffle the six, and number them.
  5. Ask a colleague or client who knows your writing to pick your three and explain each choice.

If they pick all three of yours, read their reasons; those are your fix list. Picking all three by pure guessing happens about one time in twenty, so a perfect score usually means a real difference. If they cannot reliably tell, the voice is close, but still read the AI paragraphs aloud yourself. Phrases you would never say stand out when you hear them.

As an illustration of how the feedback looks: a colleague might say she spotted the AI paragraphs because each ended with a neat summary sentence and two used the word “crucial”, which the author never uses. Both go into the “avoid” column, and the next test is closer.

Common failure modes

Failure What you see Fix
Adjective-only profile Smooth, generic prose that could be anyone’s Replace adjectives with observable traits and examples
Copied tics Your signature phrase in every paragraph List signature phrases and allow each at most once per piece
Drift over long documents The opening sounds right; later sections slide into generic prose Work in shorter sections, restate the profile, check each section
Generic AI phrasing “It’s not just X, it’s Y”, lists of three, “delve”, neat summary endings Add these patterns to the avoid column and edit them out
Invented specifics Stories, numbers or quotes you never gave it Supply the facts, forbid additions, and fact-check the result
One context everywhere Emails start sounding like posts Use samples from several contexts and name the context in each task

Why a whole book is harder than a post

A post is short enough to check in one read. A book is not, and voice problems compound across chapters.

  • Drift. Each chapter drafted in a separate session can move a little further from the profile, and the reader feels it by the middle of the book.
  • Repetition. The same phrase, example or story turns up in several chapters, because no single session sees the whole manuscript.
  • Range. A book mixes stories, explanations, instructions and arguments, and a profile built from posts may cover only one or two of those.
  • Short-form habits. Samples drawn from social posts carry hooks and one-line paragraphs into chapters where they do not belong.

What helps: use the same profile for every chapter, read chapters in order rather than one at a time, keep a running list of phrases and stories already used, and do a dedicated voice pass when you edit the manuscript. A general chat assistant can do much of this with discipline; the comparison of a chat assistant and a book platform covers where the trade-offs fall.

Privacy: check what happens to your samples

Voice samples often include client emails, proposals and unpublished drafts. Before you upload them anywhere, check:

  • Whether the tool may use your content to train or improve its models, and whether you can opt out.
  • Whether your plan’s terms differ from a personal or free plan; business plans often have different terms.
  • How long content is kept, and whether others in a shared workspace can see it.

Remove client names and confidential details from samples first, whatever the tool’s terms say.

How WritebookAI helps

WritebookAI builds this kind of voice profile for you as the first of four stages in writing one nonfiction business book. Voice builds your Writing DNA from a few samples you already wrote or said, such as posts, emails, articles and talk transcripts. It measures 60 characteristics across nine domains, from word choice and sentence shape to rhythm, rhetoric and personal voice, and shows where your samples are too thin to judge. You test it on a new topic and approve it before it writes, and one account can keep several saved voice profiles. Collect gathers your transcripts, posts and notes and shows what they cover and what is missing. Interview asks only where that material is missing, unclear or contradictory, and anything you skip is left out, never invented. Write drafts every chapter from your material in your Writing DNA, checks its factual claims against that material, and gives each chapter a Revise button for one-sentence changes. Your content is sent to AI model providers only to provide the service; WritebookAI does not sell it or use it to train AI models.

It is not a general writing assistant or a done-for-you service. It includes no copyediting, proofreading, cover design or publishing, and it cannot guarantee clients. You can start your book free with no card and read the first 3 chapters. The complete book project is $497 one time, a founding price for the first 100 authors (planned price $997), not a subscription. If within 30 days of purchase the book does not read in your voice, the “Sounds Like You” guarantee means WritebookAI works with you to refine it and refunds you in full if it cannot; see pricing for details.

Frequently asked questions

Can I train ChatGPT to write like me?

Not in the sense of retraining the model itself, which is not something a regular user does. What you can do is give it good samples of your writing and a written description of your voice, then test the output. Most assistants let you save standing instructions or reference files so you do not paste them every time; check what your plan offers, because these features change. The samples and the profile matter more than the tool, and the test tells you whether it worked.

How many writing samples does AI need to copy my style?

There is no fixed number, but several thousand words in total is a sensible guide. Variety matters more than volume: include something explanatory, something persuasive, a story and, if you can, a transcript of you speaking. Every sample should be written without AI and for real readers. A few strong, varied samples beat a large pile of similar ones, and pieces heavily edited by someone else will teach the AI their voice instead of yours.

Is it better to describe my voice or just give examples?

Use both. Samples show the voice, but an AI given samples alone tends to copy surface features, such as a favorite phrase, and miss the patterns underneath. A description alone usually turns into adjectives like warm and authoritative, which give the AI nothing specific to do. A voice profile that names observable traits, points to evidence in your samples and lists what to avoid, used together with two or three samples, gives the most consistent results.

How can I tell if AI writing really sounds like me?

Run a blind test. Pick a topic you have not written about, write three paragraphs yourself, have the AI write three from the same notes, remove formatting differences and shuffle them. Ask a colleague or client who knows your writing to pick yours and explain their choices. Their reasons are your fix list. Reading the AI paragraphs aloud also helps, because phrases you would never say tend to stand out when you hear them in your own voice.

Will an AI tool use my writing samples to train its models?

It depends on the tool and your plan. Some services may use your content to improve their models unless you opt out, and business plans often have different terms from personal ones. Read the data and privacy settings before uploading anything, especially client emails or unpublished work. Remove names and confidential details from samples regardless. WritebookAI sends content to AI model providers only to provide the service, and does not sell it or use it to train AI models.

Do I have to disclose that AI wrote in my voice?

It depends on where you publish. For books on Amazon KDP, text created by an AI tool counts as AI-generated even if you substantially edit it afterwards, and must be disclosed; AI-assisted content, meaning your own work edited or brainstormed with AI, does not need disclosure. Other platforms, publications and clients may have their own rules, so check each one. A voice that sounds like you does not change how the text was produced.

Sources

  • Amazon KDP Help: Content Guidelines: Text, images or translations created by an AI tool count as AI-generated content, even if substantially edited afterwards, and must be disclosed. AI-assisted content (your own work edited or brainstormed with AI) does not need disclosure.

Continue the book-building workflow

Turn your expertise into a book that sounds like you.

WritebookAI interviews you, organizes your talks, client stories, and notes, and drafts in your Writing DNA. Your first 3 chapters are free, then $497 one-time to finish the book.

Start your book free

See pricing