The question “How many books are written with AI?” sounds simple until you ask what counts as a book, what counts as written with AI, and who could see the whole market. No one can inspect every manuscript or reliably infer a writing process from a finished file. The useful response is a source-by-source evidence map: a study of Amazon releases, a survey of authors, a platform’s own production logs, and the rules that govern disclosure. Each has a different denominator.
This guide is an independent synthesis of public research available in September 2026. It is not a census of EarnDraft books. We do not publish an internal manuscript count here because we have not assembled and audited a suitable dataset. Where a statistic comes from one company, it describes that company’s users. Where it comes from a survey, it describes the people who answered. That distinction is the most valuable statistic to carry away.
The short answer: seven numbers worth understanding
| Evidence | Reported result | Population and date | What it does not show |
|---|---|---|---|
| NBER study of newly released e-books sold on Amazon | AI-containing titles topped half of 2025 releases | Amazon e-book releases in the researchers’ dataset; working paper published 2026 | Share of all books worldwide, or how much of each book AI wrote |
| Authors Guild author income survey | 5,699 published authors participated | U.S.-focused survey reported 2023, covering 2022 earnings | A random census of every author or an AI adoption rate |
| Authors Guild median book income | $2,000 for all respondents | Same 2023 survey, 2022 income | A forecast for a new AI-assisted title |
| Authors Guild median for full-time self-published authors | $12,800 from books | Same survey and income year | Typical income of a part-time beginner |
| Inkfluence AI’s disclosed platform corpus | 102,854 projects as of 15 September 2026 | One vendor’s production records, as described by that vendor | Market-wide output or independent verification |
| Inkfluence AI’s recorded median word count | 9,337 words | 12,647 of its projects with a stored word-count field | Median across all 102,854 projects |
| Inkfluence AI’s export signal | 40.8% exported at least once | Its full reported corpus as of 15 September 2026 | Publication, a sale, or a finished manuscript |
The NBER working paper, Authors Guild survey, and Inkfluence AI source page are separate evidence streams. We include the competitor figures with attribution because they illuminate a concrete point about measurement, not because EarnDraft can validate their private database. See our self-publishing statistics guide for the separate question of author income and publishing output.
What does “AI-written” mean?
A writer might use AI to explore an audience, generate a chapter outline, reorganize an existing draft, suggest a sentence, create an illustration, or draft every paragraph. Putting all six in one bucket destroys the information readers actually need. It also obscures the human work that matters for copyright and reader trust. A more useful taxonomy separates where AI enters the process and who makes the expressive decisions.
| Workflow | Typical artifact | Primary question for the editor | Useful disclosure record |
|---|---|---|---|
| Research assistance | Topic map or interview questions | Are the suggested facts verified independently? | Sources consulted and fact-check notes |
| Structural assistance | Outline or chapter order | Does the sequence fit a real reader’s task? | Outline versions and rationale |
| Language assistance | Rephrased human draft | Does the revision preserve the author’s meaning and voice? | Original and edited passages |
| Draft generation | AI-produced prose | Which claims, examples, and sentences has a human checked and rewritten? | Prompt, output, edits, and reviewer notes |
| Image generation | Cover or interior art | Are rights, consistency, and platform disclosures resolved? | Tool, prompt, license, edits, and asset list |
| Translation | Translated edition | Has a competent reader checked meaning, idiom, and cultural context? | Source text, translation, review record |
The U.S. Copyright Office’s AI report does not treat a mere prompt as sufficient human authorship of whatever a model returns. Human selection, arrangement, and modification can matter, and AI assistance does not automatically disqualify a larger human-authored work. Those principles are about copyright eligibility, not quality. A weak factual guide may be human-authored; a useful guide can contain AI assistance. For practical decisions, document your contribution and your checks instead of choosing an inflated label.
A concrete example. Two authors both say they “wrote a book with AI.” Mara interviewed twelve local café owners, built an outline from her notes, drafted original chapters, and used a model to tighten headings. Leon gave a broad prompt, accepted the draft unchanged, and uploaded it. Their workflow, evidence trail, expected revision load, and potentially protectable expression differ enormously. A count of “AI books” that includes both without a definition cannot guide a publishing decision.
What does the 2026 research say about book output?
The 2026 NBER working paper by Imke Reimers and Joel Waldfogel studies how large language models relate to the quantity and use of newly released e-books. Its abstract reports that the share of AI-containing books exceeded half of 2025 releases in the studied market. It also reports that AI-containing books receive less use. That pairing matters. More titles do not imply more reader attention per title. The count of publications and the value readers get are different outcomes.
“AI-containing” is deliberately broader than “entirely written by AI.” The paper’s identification method and Amazon-oriented dataset deserve attention before anyone turns its result into “half of every book is AI-written.” A book sold on a specific retailer is not a sample of a library’s backlist, print-only publications, private training manuals, or direct-sold client guides. New-release share is not the share of all books in circulation. A title count is not a page count, a sales count, or a share of reading time.
| If someone says… | Ask… | Why the answer changes interpretation |
|---|---|---|
| “Half of books use AI” | Half of which books, published where and when? | A retailer’s new e-book titles do not cover all formats and channels. |
| “AI books are growing” | Growing in listings, sales, or readership? | Supply can grow while demand per title falls. |
| “AI makes authors productive” | Productive by drafts, exports, published titles, or satisfied readers? | Each metric has a different failure mode. |
| “Authors earn more with AI” | What author group, income source, and comparison period? | Selection effects dominate casual comparisons. |
| “AI books are short” | Median of which stored manuscripts? | Missing length fields can skew the answer. |
The practical implication for an independent creator is sober: a faster draft increases supply, so the differentiator becomes the author’s firsthand knowledge, specific examples, careful editing, and distribution. Our ethical AI business book guide shows a workable review process; the book for consultants guide shows why a narrow audience can be more useful than a broad topic.
Can any source count every AI-written book?
No current public source can observe every manuscript’s production history. Retailers can see some disclosures and files they receive. Tool companies can see activity inside their own systems. Researchers can sample retail listings and infer likely use with uncertainty. Surveys can ask authors, but the answer depends on memory, definitions, and willingness to disclose. A publishing institution may count registered titles, but a title record rarely says how each sentence was composed.
Consider a simple coverage model. Let A be books drafted inside a particular tool, B books released through a retailer, and C books whose writers answer a survey. These sets overlap, but none contains the whole market. A writer may draft in one tool, finish in a word processor, sell a PDF directly, and never answer an industry survey. Another may generate an outline in a chat tool and publish through three stores. Adding A, B, and C would count some works more than once while missing others entirely.
The only defensible total is a clearly bounded one: “projects created on platform X between these dates, under this definition.” That is why the Inkfluence corpus is useful as a platform case study and unsuitable as a global estimate. If EarnDraft later publishes an audited dataset, the page should state inclusion rules, exclusions, observation dates, missing fields, and what “finished” means. Until then, we prefer to show readers how to evaluate evidence rather than invent a dashboard.
How should you read a platform’s book statistics?
Inkfluence AI reports 102,854 projects created through 15 September 2026. Its own methodology says the figure includes unfinished projects and excludes the team’s test books. A project is not necessarily a book someone has read or bought. The source page also reports that only 12,647 projects carried a stored word count; its length median therefore uses that subset. The result may be helpful, but missingness matters because newer or more actively edited projects may be overrepresented. These are the kinds of caveats a good statistics page should place next to the number, not in invisible footnotes.
Its reported export rate of 40.8% is an activity signal. Some authors export to proofread; some finish inside an editor and do not export; some export several times. Export does not imply publication. Listing through a store does not imply a sale. A sale does not imply the buyer finished reading. The funnel can be represented as project → draft → edited manuscript → export → publication → discovery → purchase → completion, with leakage at every step. If a vendor reports only the left side of the funnel, it cannot tell you the right side.
A denominator exercise
Imagine a tool reports 10,000 projects and 4,000 exports. “40% exported” is an accurate arithmetic statement under a precise definition. If 2,000 projects are duplicates, tests, or abandoned starts, a different denominator changes the interpretation. If 500 exports are early drafts, “40% finished” is an overclaim. If 300 titles reach a storefront, “40% published” is wrong. The exercise does not challenge any particular vendor’s records; it shows why nouns in a statistic matter as much as percentages.
| Funnel stage | Evidence a system might hold | What it cannot infer automatically |
|---|---|---|
| Project created | Timestamp, account, chosen format | Whether meaningful writing began |
| Draft generated | Chapter files, revision events | Whether the text is accurate or useful |
| Exported | File type and timestamp | Whether the author considers it final |
| Listed | Store integration event or public URL | Whether a buyer paid |
| Sold | Transaction data from a specific channel | Sales through other channels |
| Read | Reader data or voluntary feedback | Satisfaction for all readers |
For a creator, the relevant metric is often further right: did the book help a reader complete the promised task? A consultant may benefit from qualified calls after a free guide, with no direct book revenue. A workbook seller may care about completion and repeat purchases. A novelist may care about read-through to book two. Pick the outcome first, then choose a metric that measures it.
Are AI-assisted books mostly fiction or nonfiction?
One vendor’s genre mix is not a market mix. Inkfluence reports that 63.7% of its classified projects were nonfiction and 36.3% fiction. That is an interesting description of its templates, positioning, and customer base. It cannot prove that global AI-assisted publishing has the same split. The platform offers specific blueprints; categories made easy to create can become common in its logs. The categories themselves may be assigned from the selected workflow rather than an independent content review.
The point still generates a practical question: what is the job of this book? A five-chapter client guide, a thirty-day exercise book, and a romance manuscript require different editorial standards. A short guide is not a failed novel. A long novel is not a better lead magnet because it has more words. Match length, visual support, and delivery format to the reader’s task.
| Reader task | Suitable content shape | Evidence of usefulness | Editing risk |
|---|---|---|---|
| Decide whether to hire an expert | Narrow guide with decision criteria and examples | Qualified questions, useful referrals | Vague claims and hidden sales pitch |
| Learn a repeatable process | Workbook with steps, exercises, answer space | Exercises completed, feedback on confusing steps | Instructions that cannot be followed |
| Research a destination | Travel guide with verified logistics and maps | Correct opening hours, route choices, recent updates | Stale prices and imagined venues |
| Enjoy a story | Fiction with coherent character, scene, and pacing | Reviews, read-through, editorial feedback | Inconsistency and generic prose |
| Prepare for a purchase | Buyer’s guide with comparisons | Better decisions and fewer surprises | Affiliate bias and inaccurate specifications |
EarnDraft is best used as a starting point for a human-led book workflow. Start with a focused brief, inspect the outline, then verify every claim and add your own examples. The guide hub covers specific publishing choices, while creating a book begins the draft flow. Do not represent these links as proof of any particular export or publishing integration; check the current product capabilities before promising a file format.
What does book length tell us, and what does it hide?
Word count has one virtue: it is easy to measure when a manuscript file is available. It is also a poor proxy for usefulness. A 3,000-word checklist with original examples may solve a problem better than a 30,000-word general survey. Yet length matters for reader expectation and format. A customer buying a full travel guide expects depth; a lead magnet promising a quick answer benefits from restraint. The proper comparison is with the promise on the cover and sales page.
Inkfluence reports a median of 9,337 words among the 12,647 projects with stored counts, not among its entire 102,854-project corpus. If missing counts are correlated with project age, genre, or completion, the full-platform median could differ. A median also hides the long tail. If half the books are under 9,337 words and half above, it says nothing about the shape of those halves or the quality of any manuscript.
For an author planning a book, use a chapter budget rather than a market median. Write a one-sentence promise, list the decisions or tasks needed to fulfill it, and estimate the space each requires. Then produce one representative chapter before locking the target length. If your model chapter uses 900 words plus a diagram and an exercise, six chapters may make a substantial 6,000-word asset. If it uses case studies and detailed calculations, it may need far more. The chapter budget is a planning hypothesis, not a quota.
A practical length worksheet
| Book component | Example purpose | Draft allowance | Cut or expand based on… |
|---|---|---|---|
| Orientation | State the reader and outcome | 400–700 words | Whether a visitor can decide this is for them |
| Core chapter | Teach one decision or task | 800–1,800 words | Whether examples answer actual questions |
| Worked example | Show the method in context | 300–800 words | Whether readers can apply the result |
| Reference table | Make comparison skimmable | As needed | Whether the table simplifies a real choice |
| Exercise | Turn reading into action | 150–500 words | Whether the prompt has an observable output |
| Closing plan | Tell readers what to do next | 300–600 words | Whether next steps are concrete |
These ranges are planning examples, not measured publishing norms. They are intentionally broad because a visual workbook and a technical reference do different work. Test the book with a reader before adding words to meet an arbitrary threshold.
Do exports reveal completion?
Only partially. “Exported at least once” is a reasonable operational metric for a writing tool because an export usually means a user wanted a portable copy. It is not a verified finish event. A proof PDF can be exported after chapter one; a polished book might remain in a connected publishing system without an export. Some platforms count the most recent export format, which can obscure earlier formats. The more precise sentence is: “X% of projects had an export event under these rules.”
Authors can design a better completion rubric for their own work. First, a complete manuscript has all promised sections. Second, every factual claim has a source or documented firsthand basis. Third, a reader unfamiliar with the topic can use the examples. Fourth, the file passes device and print checks. Fifth, the title and sales description accurately reflect the contents. Only then call it ready. Publishing is a separate event; commercial success is another.
The distinction matters for SEO and marketing too. A public “books created” counter draws attention but tells a buyer little about results. Case studies with a real audience, a described editorial process, a sample page, and an outcome answer a more valuable question. If you are evaluating any AI book tool, look for the route from draft to reviewed artifact, not only the speed of the first draft.
What can author-income data tell an AI writer?
The Authors Guild survey is among the best-known public sources on U.S. author income, but it was not designed to estimate how much an AI-assisted book earns. It surveyed 5,699 published authors about 2022 income and reported a $2,000 median book income across all respondents, including part-time authors. Full-time self-published respondents reported a $12,800 median from books. Those are group medians, not guarantees, and the groups differ in experience, catalog size, genre, marketing, and time invested.
The survey also separates book income from other author-related work such as teaching, speaking, editing, or coaching. That is especially important for a creator making a business book. A guide might earn little in direct royalties and still support a valuable service business. Counting only royalties would miss the outcome; attributing every client to the book would overstate it. Track a simple source question in inquiries, then compare with a baseline.
| Question | Better measure | Weak shortcut to avoid |
|---|---|---|
| Does the book sell? | Net revenue after platform fees, refunds, and ads | Gross list price multiplied by downloads |
| Does it bring qualified leads? | Calls or subscribers who cite the book and fit the offer | Total website visits |
| Does it build trust? | Feedback on specific ideas, invitations, referrals | Social impressions alone |
| Is it worth updating? | Continuing use relative to maintenance hours | First-week launch spike |
| Is a series working? | Reader progression from one title to the next | Number of titles published |
For royalty mechanics, see KDP ebook royalties. For selling a practical file rather than a retailer ebook, see how to sell workbooks. Neither guide should be read as a forecast of personal earnings.
Does more AI content mean more reader demand?
No. The NBER paper’s distinction between quantity and usage is a reminder that content supply can grow faster than attention. Readers have limited time. Retail search has limited first-page real estate. A larger catalog raises the importance of discoverability, trust, and genuine usefulness. Publishing more pages with weak evidence can also harm a brand’s credibility even if each page is indexable.
For creators, this changes the editorial question from “How quickly can I generate a book?” to “What unique information can I make dependable?” Interviews, field observations, annotated examples, original diagrams, and careful synthesis are harder to substitute. A real instructor can explain which exercise students fail and why. A local host can verify whether a route remains accessible. A consultant can show how a decision changed in a client case after securing permission and removing private information.
An AI draft can help organize that material, but it cannot manufacture firsthand evidence. Treat a model’s confident sentence as a hypothesis. Put citations near claims. Mark dates for prices and policies. Maintain a revision log. Let your book earn its space through specific utility rather than the fact that a publishing tool made it possible.
How to evaluate an AI book statistic in sixty seconds
Use the six-question check below when a chart, social post, or vendor landing page cites a large number. You do not need a statistics degree; you need the denominator and the verb.
- Who counted it? A publisher, retailer, researcher, survey firm, or tool vendor? Each sees a different slice.
- What exactly was counted? Accounts, projects, chapters, exports, ISBNs, store listings, sales, or reads?
- When and where? A date range and geography matter; “2026” alone is insufficient.
- How much is missing? If only a subset has word counts, compare the subset with the whole.
- What is the comparison? A change from 2% to 4% is a doubling and a two-point increase. Both can be true.
- Could the claim survive a weaker verb? “Exported” is observable; “finished” may be an inference. “Listed” is observable; “earned” requires transactions.
Practice exercise: rewrite a headline
Suppose you read “AI writers finish forty percent of books.” Rewrite it as “On one platform, forty percent of book projects had at least one export event during the measured period.” Then add the source, the count date, and the definition of a project. The revised version sounds less dramatic and helps readers make a better decision. Do this exercise with any impressive number before repeating it in your marketing.
What evidence should a creator collect for their own book?
A solo author does not need a research department. A modest evidence file prevents many avoidable errors. Save the original source for each numerical claim, the date you checked it, and a sentence about what it actually supports. Record the intended audience and promise. Keep a small sample of reader feedback, including confusion and criticism. Track revision dates. If the book gives practical instructions, run through them yourself or ask a qualified reviewer to do so.
| Evidence artifact | Minimum useful record | Reason to keep it |
|---|---|---|
| Claim ledger | Claim, source URL, access date, page or section | Makes later updates efficient |
| Editorial log | Draft version, reviewer, decisions | Shows which judgments were human |
| Permission log | Image, quote, testimonial, and use rights | Reduces rights ambiguity |
| Test record | Device, file format, page or chapter, issue | Prevents broken reader experience |
| Update note | What changed and why | Signals maintenance to readers |
For AI-assisted work, add the model or tool used where it affects rights or disclosure, plus what the author changed. Do not publish sensitive prompts or client data merely to prove involvement. The goal is a truthful record that supports quality and compliance.
Methodology and limitations of this page
We reviewed the linked primary materials and extracted only claims they could support as of September 2026. We did not run a survey, access a retailer’s private sales records, or audit another tool’s database. The NBER result comes from a 2026 working paper and should be interpreted within its studied channel. The Authors Guild income figures come from a survey of respondents, not an exhaustive tax record. Inkfluence AI’s project, genre, length, and export figures are its own disclosed measurements; we attribute them and preserve its stated missing-data caveats.
This page is intended as a method for reading the market, not as a market-size forecast. Definitions and platform policies can change. Revisit each source before relying on a figure in an investor deck, legal filing, or public advertisement. If an updated dataset appears, keep the old count date visible instead of quietly changing the headline.
Frequently asked questions
How many AI-written books exist worldwide?
No verified worldwide count exists. Studies can estimate AI-containing titles in a defined retailer dataset, while tool companies can count projects on their own platforms. These are different denominators, and neither captures every manuscript or publication channel.
Are half of all books now written by AI?
That wording overstates the evidence. The NBER working paper reports that AI-containing books topped half of newly released e-books in the market it studied in 2025. “AI-containing,” “newly released,” and the retailer dataset are all essential qualifications.
What is the average length of an AI book?
There is no reliable global average. Inkfluence AI reports a 9,337-word median among 12,647 projects with stored word counts on its platform as of 15 September 2026. The result does not establish a median for all AI-assisted books or even every project in that vendor’s corpus.
Does an export mean the book was finished?
No. An export is a file event. An author might export a draft for review, or finish a project without exporting. It is a useful operational proxy only when labeled clearly.
Can I cite a vendor’s private platform statistics?
Yes, if the vendor publishes them for reuse under its terms and you attribute the source, count date, denominator, and limitations. Do not present them as independently verified industry totals. Inkfluence AI describes its public CSV files and attribution terms on its statistics page.
Does AI assistance make a book uncopyrightable?
Not automatically. The U.S. Copyright Office describes how human-authored expression may be protected even when AI assists, while purely AI-determined expressive material may not qualify on its own. The result depends on the actual human contribution and jurisdiction; see the Copyright Office report.
Which number should I use to plan my own book?
Start with the reader promise and a representative chapter. Use a word budget to plan the work, then revise based on examples, testing, and reader feedback. A platform median is context, not a target.
How often should a statistics page be updated?
When its underlying sources change, with a visible review date and changelog. A calendar year in a title does not mean the data were recounted that year. Preserve the source’s observation date so readers can judge freshness.
Sources and further reading
- NBER, AI and the Quantity and Usage of Creative Products: Have LLMs Boosted Creation of Valuable Books?, working paper, 2026. Research context for Amazon e-book releases and usage.
- Authors Guild, 2023 Author Income Survey key takeaways, published 2023. Survey sample, income definitions, and median figures.
- U.S. Copyright Office, Copyright and Artificial Intelligence, including Part 2 on copyrightability. Primary source for U.S. human-authorship guidance.
- Inkfluence AI, AI Book Writing Statistics 2026, counted 15 September 2026. Attributed platform figures and its methodology; not independently audited here.
- Amazon KDP content guidelines, current rules for content quality and AI-generated content disclosure.
The next useful step is to define a book only a human with your experience could make: a named reader, a specific outcome, an evidence file, and a review plan. Start a focused draft, then return to these measurement questions before you publish.