Interactive eBook BuilderInteractive eBook Builder
$59 · unlimited access, unlimited eBooksGet it for $59

· Updated · By · 7 min read

How to Stop AI Making Up Facts in Your Nonfiction Ebook

aipromptsebookpublishing

You can't fully stop AI from making up facts, but you can sharply reduce the risk by giving it your sources, restricting it to them, requiring a supporting quote for every claim, and checking each number, quote and date yourself before you publish. Hallucination is a known property of language models, not a bug you can prompt away.

This guide gives you a source-first workflow, three prompt patterns from the model makers' own documentation, a claim ledger you can copy, and a short list of what to double-check last.

Why does AI make up facts?

Because a language model generates likely-sounding text, not verified text. The problem has been studied for years: a 2022 survey of hallucination in natural language generation (Ji et al.) catalogs it across tasks, and dialogue-system research found that adding retrieval from a knowledge source substantially reduced hallucination (Shuster et al., 2021). The lesson is the same for ebook authors as for researchers: grounding the model in real documents helps.

Anthropic's guide is blunt on the limits. It lists techniques that "significantly reduce hallucinations" and then adds: "they don't eliminate them entirely. Always validate critical information." So the plan is risk reduction plus human verification, not a magic prompt.

What does a source-first workflow look like?

You collect the facts first, the AI writes second, and you verify third. Most mistakes happen when people reverse this and ask the AI to supply facts.

  1. Gather sources. Your notes, interviews, official pages, papers. Open each one yourself.
  2. Extract facts into a short list with the source next to each (see the ledger below).
  3. Prompt with the list and a "use only these" rule.
  4. Verify the draft against the ledger, and anything outside it gets removed or sourced.
  5. Disclose and publish following the platform's rules.

This is slower than "write me a chapter with statistics", and that's the point. A chapter with three verified numbers beats one with ten unverified ones. If you're writing for a niche audience, readers will notice a wrong number quickly, and refunds and reviews follow.

Which prompt patterns reduce made-up facts?

Four patterns from Anthropic's hallucination guide work in any assistant:

PatternWhat it saysWhy it helps
Allow "I don't know""If you're unsure or the notes don't say, write: I don't have enough information."Removes the pressure to fill gaps
Quote first"Extract word-for-word quotes from the sources, then base your analysis only on those." (recommended for long documents over 20k tokens)Grounds claims in actual text
Cite and retract"After drafting, find a supporting quote for each claim. If you can't, remove the claim and mark it with []."Makes the draft auditable
Restrict to provided text"Use only the documents provided, not general knowledge."Stops silent invention

Anthropic also suggests running the same prompt several times and comparing; inconsistencies may signal hallucinated details (best-of-N). A fourth idea comes from Chain-of-Verification research (Dhuliawala et al., 2023): draft, plan verification questions, answer them independently, then write the final version, and that reduced hallucinations in their tests. Here is a combined prompt:

Write the section "[heading]" (about 400 words) using ONLY the facts in <facts> below.
Rules:
- Every number, date, name and quote must come from <facts>. Do not add any others.
- If a point would be stronger with data we don't have, write [DATA NEEDED: what] and continue.
- After the draft, list each factual claim and the fact number that supports it. Remove any claim without support.
<facts>
1. [fact] (Source: [name, link])
2. [fact] (Source: [name, link])
</facts>

Remember that a model checking its own work isn't independent. Research by Huang et al. (2023) found that LLMs struggle to correct their own reasoning without external feedback, which is why the ledger check below is a human job. For the broader recovery routine, see what to do when AI gets it wrong.

What is a claim ledger?

A table listing every checkable claim in your book with its source and a verified box. It takes minutes to start and saves hours at the end. Here's the format, with a made-up example:

#Claim (as written)TypeSource (opened by me)Verified?
1"The tax filing deadline is [date]."Date[official page, retrieved date]Yes
2"62% of freelancers under-save."Statisticnone foundRemove
3"As [Name] said, '...'"Quote[book, page]Pending
4"Study X found Y."Citation[paper link]Yes (checked abstract)

Row 2 is the common case: a plausible-looking statistic with no source. If you can't find it in five minutes, delete it. Rules for what to check hardest (my ordering, so treat it as opinion): quotes attributed to real people, statistics, named studies, URLs, laws and regulations, and medical, legal or financial specifics. For anything that could affect a reader's health, money or legal position, send it to a qualified professional or write around it.

The ledger is also a natural fit for a research step in a workflow. In Interactive eBook Builder, the Packs Database includes Research Pack entries, so the verified facts for a chapter can sit next to its Section Blueprint rather than in a separate document you have to remember to paste.

What about disclosure and responsibility?

You're responsible for accuracy, and some platforms require disclosure. Amazon KDP's content guidelines say you must inform them of AI-generated text, images or translations when publishing or republishing, and they define the term: if an AI tool created the content, it's AI-generated "even if you applied substantial edits afterwards". AI-assisted content (you created it and used AI to edit or refine it, or only to brainstorm) doesn't need disclosure. The same guidelines say you're responsible for verifying that content follows all KDP content guidelines, including intellectual property rights. Details are in our post on KDP AI disclosure, and the wider question of using chatbots for books is covered in can ChatGPT help you write a book?

Because the ledger approach keeps your own research at the center, it also moves the work toward AI-assisted rather than purely AI-generated, but that classification is ultimately Amazon's, so read the current rules.

What should I check last?

Do one final pass for the claims that survive every other step: the numbers in headings and callouts, the quotes, dates, any "studies show", links, and product or tool descriptions that may be out of date. Models also describe tools and prices as they were at training time, so verify those against a live page. If you want a staged process around this, prompt chaining for planning, drafting and editing shows where a fact-check stage can sit. In Interactive eBook Builder, the final QA pass in Agent C Polisher is a convenient checkpoint for running your ledger, although it can't verify facts for you; that part stays with the author.

Before you hit send

The takeaway: facts come from your sources, wording comes from the AI, and verification comes from you.

FAQ

Can I trust AI for facts in my ebook?

Not without checking. Language models produce plausible text and can state wrong facts confidently. Treat AI as a drafting tool and verify every statistic, quote, date, name and citation against a source you've opened yourself.

What kinds of claims are most risky?

Specific, checkable ones: statistics, dates, direct quotes, named studies and URLs. A vague statement is less likely to be flagged as wrong but also less useful. In my judgment (opinion), numbers and quotes deserve the closest review.

Does telling the AI "don't make things up" work?

Only a little. More effective is giving it the facts, limiting it to them, allowing it to say it doesn't know, and asking for a supporting quote for each claim.

Do I have to disclose AI use to Amazon KDP?

Amazon's content guidelines require you to inform them of AI-generated text, images or translations when publishing or republishing, but not AI-assisted content. Check the current guidelines before you publish.

Ship niche eBooks with agents already built.

Get the Interactive eBook Builder for $59