How to Protect Your Original Ideas When Writing With AI

Three different worries travel under the same sentence. One coach means “I don’t want the book to end up saying what every other coaching book says.” A consultant means “I can’t have my client’s restructuring plan sitting on someone’s training server.” And a founder means “If AI helped write it, is it even mine?”

All three are reasonable, and they have different fixes. Most advice mixes them up, so people either freeze (no AI at all) or shrug (paste everything). Neither is necessary.

Three meanings of “protect”, three different fixes

What you are protecting	The real risk	What protects it	Who decides
Your thinking	The book drifts toward the typical view on your topic	Writing positions before prompting; using AI to argue, not to originate	You
Your data and your clients’ data	Confidential material is stored, reviewed or used for training under terms you never read	Pasting less; anonymising; the right plan and setting; client consent	You, your contracts and the provider’s terms
Your ownership	Parts of the book may not be protected by copyright	Human authorship you can show; your own drafts and edits	The law (and, separately, the tool’s terms)

Our view: the first row matters most and gets the least attention. A leaked paragraph is rare. A book that quietly argues the median position is common, and it costs you every client looking for a reason to pick you over the next coach.

Meaning one: keep your thinking yours

A language model produces likely text. Ask it for “the key principles of leadership feedback” and you get the principles most often written about. That is a useful map of the consensus and a poor source for the part of your book that should disagree with it.

What the research says about convergence

Vishakh Padmakumar and He He ran a controlled study (ICLR 2024) in which people wrote argumentative essays alone, with a base GPT-3 model, or with InstructGPT. Writing with InstructGPT, but not the base model, significantly reduced diversity: essays by different people became more alike. The drop came mainly from the model’s own contributions; the human-written parts were “unaffected by model collaboration”.

Anil Doshi and Oliver Hauser found something similar in short fiction (arXiv preprint): writers given AI story ideas produced work rated more creative, but the AI-assisted stories were more similar to each other.

The third study is the one we would put in front of every author. Maurice Jakesch and colleagues (CHI 2023) had 1,506 people write about whether social media is good for society, some with an assistant set up to favour one side. The assistant changed which views people wrote, and it “shifted their opinions in the subsequent attitude survey”. The suggestions shaped what the writer thought, not just the essay.

None of these studies used business books, and lab essays are not a 50,000-word manuscript. But the direction is consistent: when the model goes first, your writing and possibly your opinion move toward its defaults. Our article on why AI-written books sound generic covers the prose side; here we care about the ideas.

The positions-first exercise

Do this once for the whole book before you outline (about 45 minutes for ten positions), then again before each chapter. For each position, write four lines in your own words, in a document that lives outside any AI tool:

1. I believe: the claim, stated so someone could disagree with it.
2. The usual advice says: what a competent competitor or a popular book would say instead.
3. Because: the reason you think the usual advice fails, from your own work.
4. I learned it when: the client situation, project or mistake that taught you (anonymised).

A hypothetical executive coach, writing about feedback in senior teams, might produce this:

I believe: Most leadership teams do not need more feedback; they need fewer people allowed to give it. The usual advice says: Build a feedback culture where everyone gives feedback to everyone, often. Because: In teams over eight people, open feedback turns into lobbying. The useful feedback comes from two or three people with context. I learned it when: A team I coached ran monthly 360s for a year and trust fell each quarter.

Now ask a chatbot for “principles of feedback for leadership teams” and compare. If its list reproduces your position, either your position is the consensus (then it is not your differentiator) or you have not stated it sharply enough. If the list contradicts you, good: that contradiction is the chapter.

Date the file and keep it. It records what you thought before any tool touched the book and doubles as the brief for every chapter. If you struggle to fill ten positions, the problem is earlier than AI; start with extracting your knowledge before writing.

Then let the AI argue, not originate

Once your positions exist, the prompts change from “give me ideas” to:

• “Here is my position. Give me the three strongest objections a sceptical CFO would raise.”
• “What cases would contradict this? Which kinds of client would it fail for?”
• “Draft this section from my notes below. Do not add claims, examples or statistics that are not in the notes.”

You decide what survives each round. That judgement is the original part. For drafting habits that stop the model adding things you never said, see how to stop AI inventing facts in your book.

Meaning two: what happens to what you paste

Most anxiety sits here, and the facts change often, so everything below is dated.

Consumer plans and business plans are not the same product

As of November 2026, Anthropic’s consumer terms update, announced on 28 August 2025, applies to Claude Free, Pro and Max (including Claude Code used from those accounts). Users choose whether their chats may be used to improve the models; existing users had until 8 October 2025 to choose. With the setting on, data from new or resumed chats is retained for five years; with it off, the existing 30-day retention applies. You can change it in Privacy Settings at any time, though data already in a training run stays there. The update does not apply to Claude for Work (Team and Enterprise), Claude for Government, Claude for Education or API use, including through Amazon Bedrock and Google Cloud’s Vertex AI, which run under commercial terms.

OpenAI’s help and policy pages on how ChatGPT data is used for training returned an access error (HTTP 403) when we tried to open them for this article, so we will not state OpenAI’s current settings from memory. Check your own ChatGPT settings and OpenAI’s official help page on the day you decide.

Client material is a different category

Your own unpublished framework is yours to risk. A client’s information is not.

The ICF Code of Ethics, in effect since 1 April 2025, defines confidentiality as the protection of “any information obtained in or around the coaching relationship”, with exceptions for legal requirements, risk of harm or the client’s written consent. Standard 2.1 asks for “the strictest level of confidentiality with all parties involved”. Standard 2.5 extends a coach’s obligations to any technology they use, and names software and artificial intelligence among the examples. In other words, pasting a session transcript into a chatbot is an ethics question before it is a privacy-settings question.

Consultants have NDAs that often say more than any code. If you are unsure what you may say about a client at all, read how to use client stories and case studies in a book before you open any AI tool; the craft side is in our main-site guide on using client stories in a business book.

Our line: if a passage about a client was drafted with AI help, say so in the permission request, and never upload their recordings or transcripts to a general chat tool.

Meaning three: who owns the words

Ownership has two layers. The first is contractual: does the provider claim anything in its output? Anthropic’s consumer terms, effective 8 October 2025, say you retain your rights in the inputs you submit, and that Anthropic assigns to you “all of our right, title, and interest—if any—in Outputs”. Note the “if any”: the provider is passing on whatever it has, not promising that copyright exists.

The second layer is legal, and it decides whether anyone can claim copyright at all. On 29 January 2025 the US Copyright Office released Part 2 of its AI report. According to the Library of Congress announcement, AI output “can be protected by copyright only where a human author has determined sufficient expressive elements”, which can include a human-authored work that is “perceptible in an AI output” or a human making “creative arrangements or modifications of the output”, “but not the mere provision of prompts”. Using AI to assist, or including AI material in a larger human work, “does not bar copyrightability”.

For a business author, the practical reading is reassuring:

• Passages you wrote yourself are yours in the ordinary way.
• AI-drafted passages built from your own writing, with your structure and your edits, have a much stronger claim to human authorship than text produced from a one-line prompt.
• The Copyright Office’s own FAQ says copyright “does not protect ideas, concepts, systems, or methods of doing something”, only how you express them. In practice, what protects a framework is publishing it well under your name.

Keep your positions document, notes and a few dated drafts as a record of what you contributed. This is the US position only, other countries differ, and if a licence or publishing contract depends on it, ask a lawyer. Nothing here is legal advice.

Where guided software fits

General chat tools start from your prompt; guided software can start from your material. WritebookAI is one option built that way: you upload PDF, Word or text files, its source analysis shows what your material covers and what is missing, and a verification interview asks you about gaps and conflicts. Chapters are drafted from your material, and when something is missing, such as a client result, the chapter leaves it out and asks you rather than inventing it; you approve every chapter. You own the book and keep 100% of royalties (the contractual layer, not a copyright guarantee). It does not decide what is safe to upload or anonymise client material for you. Read its privacy policy before uploading anything sensitive, and see the WritebookAI homepage for what is free to start. If you would rather stay in a general tool, the method in our guide to writing a book with AI from your documents works there too.

Three mistakes that make this harder than it needs to be

Refusing AI entirely. With your positions written and client material anonymised, the remaining risk to your own unpublished ideas is modest. Skipping AI to protect an idea that will be in print within a year is usually the wrong trade.

Treating a privacy setting as a confidentiality policy. Turning off training does not give you the right to share a client’s information. Consent and anonymisation come first; the setting comes second.

Asking the model for your chapter list. The outline is where your argument lives, so write it from your positions and ask the model only what is missing. If you do not yet have a named model of your own, creating a framework for your book is the better starting point.

A one-page record to keep for every book

Before you write another chapter, open a single document with three headings: Positions (your ten four-line positions, dated), What I share with AI (which account, which plan, the training setting and the date you checked, plus your rule for client material) and What I wrote (where your drafts and edit history live).

Fill the first section this week; it alone makes the book better as well as safer. The other two take ten minutes, and if a client or publisher ever asks, the answer is on one page.