The endgame of publishing’s AI-generated book problem

There have now been at least two books whose publishing deals were killed because of accusations of AI-generated content. Mia Ballard’s novel Shy Girl was published and then withdrawn by Hachette Group and Jerry Falade’s crime novel Call Me, I’ll Hide the Body received a $2 million offer after an auction, and then was withdrawn by Falade’s agent. Ballard says AI content was inserted by her editor and Falade denies using AI to write his book.

Anyone who hasn’t dealt with AI-generated text might look at these examples and say “It’s simple. Don’t publish AI-generated text. Don’t pay advances for it, don’t print it, and don’t carry it in bookstores.”

It’s not simple. It’s not as if AI-generated text flashes in purple all-caps. Publishers are not stupid.

I can tell you where this is going, not just based on my close observation of the industry, but on what has now become extensive experience dealing with authors and AI, and with my own personal experience with AI.

The hard facts

AI use by authors and publishers is an emotional issue. Put the emotions aside for a moment. These are facts:

• AI-generated text is not eligible for copyright. The US Copyright Office’s statement on this issue is unequivocal. When you submit a work for copyright registration, you must indicate if any parts of it are AI-generated, and any such parts are not protected by copyright. It doesn’t matter how clever your prompts are: if the text is output from AI, it’s not eligible for copyright.
• Major publishers won’t publish work they know is AI-generated. Publishing deals (including both traditional deals and hybrid publishing deals) require copyrightable text. The publisher depends an exclusive license to publish the text in the book; copyright is what makes that exclusive license possible. The publisher will commit significant editorial, production, printing, and financial resources to publishing a book; without that exclusive license, anybody else can publish the same book, so the publisher will reject it.
• Many authors use AI. My 2025 survey with Gotham Ghostwriters, “AI and the Writing Profession,” reached hundreds of authors. In that survey, 58% of nonfiction authors, 42% of fiction authors, and 68% of ghostwriters said they used AI. They were most likely to use AI to replace search, to replace a thesaurus, for other research, and to generate text that they then edited further. Using AI to write was rare but not unheard-of: 7% of nonfiction authors, 11% of fiction authors, and 5% of ghostwriters admitted generating publishable text with AI.
• There is no dependable test for AI-generated content. The best available AI-testing software is Pangram. If Pangram says a book is human authored, it probably is. But if Pangram says parts of a book are AI-generated, it can be in error (that is, there can be false positives). Pangram states it only generates a false positive one time in ten thousand, but I don’t buy it. I have personal experience with this: I’ve seen it accurately mark both AI-generated text and human-authored text. I’ve also extensively revised AI-generated text and seen Pangram still mark it as AI-generated.
• AI “tells” are constantly shifting. Editors and intelligent readers will tell you that AI-generated text is like Supreme Court Justice Potter Stewart’s view of pornography, “I know it when I see it.” If this is you, I promise you that you are wrong. AI used to use lots of em dashes; now it doesn’t, and plenty of human authors do. AI generated text tends to embrace the “rule of three,” as does lots of human-created text. AI text is full of “It’s not this, it’s that,” constructions, at least for now, but there’s no reason to believe that will be true of whatever Large Language Model versions are released in a month or two. Any “definitive” list of tells is useless, because humans often use the same constructions, and there’s no assurance that models won’t cure their verbal tics in the near future.
• Human-authored certifications are flawed. You can get a badge that says your work is human authored simply by promising that you didn’t use AI to write it — but should we believe you? The Authors Guild charges a minimal fee for you to certify that your work is human-written, which assumes the author isn’t lying or deceiving themselves. You can also pay to have your work tested by VerifyMyWriting; this site tests your work with Pangram and issues a verifiable certificate if most of it checks out to be human. Because of the weaknesses of Pangram, failing this certification doesn’t prove that you used AI to write your book.
• Publishers are understaffed. Editorial staff at publishers continue to be overwhelmed. Publisher typically expect authors to do most of their own editorial work, or hire a developmental editor. In this environment, editorial staff at publishers don’t have time to conduct a detailed analysis of whether and how a work might be (or seem) AI-generated. If you’re thinking “publishers should just know,” you’re being unrealistic about the work required.

The endgame

Harper Collins CEO Brian Murray recently told Publishers Weekly that issues regarding AI in publishing “need industrywide solutions.” That will take years, if it ever actually happens. In the meantime, the industry is going to evolve.

How? Its time to put my analyst hat on and predict the future. (Predicting the future of media industries was my job for 20 years — this is familiar territory for me.)

Here’s what I think will happen.

• AI text-generation will get better and better. Models will improve; tells, such as they are, will become rarer and rarer. AI-generated text will become increasingly harder to detect, both by humans and by machines.
• AI-detection software will get better, but will never catch up. Each generation of AI models will incorporate more features to reduce such tells. The AI detection software generally analyzes how text follows predictable patterns or not (a quality called “perplexity”). Models will develop that are specifically designed to evade such analysis.
• Publishers will count on a combination of AI-testing, author assertions, and instinct. To get a publishing contract, you’ll need to attest that your manuscript does not include any AI-generated text, with limited and specific exceptions. The publisher will run your manuscript or proposal through a tool like Pangram, and they’ll discuss with you how you created the manuscript. In the end, if your manuscript seems fully original and fails to set off any alarm bells, and you sign a document that you didn’t use AI to generate it, they’ll accept that. The point will not be whether it was AI-generated, but whether it was obviously so full of AI that the copyright office and readers would reject it. The witch hunts will subside.
• Manuscript quality and marketability will continue to matter most. Right now, you need three things to succeed with a book: a great idea, excellent writing, and a solid promotional plan. That is not going to change because of AI. It will never substitute for the human qualities that make a book great. That writing quality, plus an excellent promotional platform, is what will make you (and your publisher) successful.
• Authors guidelines will emerge. We already have solid guidelines for ghostwriters. The Authors Guild has excellent guidelines as well. These sorts of guidelines will coalesce into evolving best practices.
• Authors will get good at using AI for the right purposes. Authors who start with a human idea and natural writing talent will succeed. AI will help them do faster research, stress-test their ideas, identify problematic writing tendencies, and suggest places where ideas were inadvertently plagiarized.
• Self-published books will be full of AI. There’s no effective way to police this. There have always been loads of crappy self-published books. Now there will be a new thick layer of slop below the crap. Amazon will create policies to put the brakes on (as it did when it limited the number of books a person could post in a day) but those policies won’t meaningfully stop the flow of slop. Book buyers and readers will react by increasingly making choices based on recommendations from friends and other trusted reviewers.
• Readers will focus on entertainment and value, not AI. Accusations of AI use will matter less and less. Fiction readers want entertainment. Nonfiction readers want useful advice. Whether a few paragraphs were spewed out by an AI tool will matter less than whether the book is actually any good.

Bottom line: AI won’t kill the publishing business. It won’t significantly change how readers, bookstores, or publishers behave. Good books have always sold and always will.

The onus is on authors to 1) learn to use AI in responsible ways, not for writing text, 2) expect to be able to document your writing process, and 3) focus, not on text generation, but on becoming a great writer with great ideas.

And if you’re waiting around for Brian Murray’s “industrywide solutions,” you could be waiting a very long time.