In March, a New York Times investigation found evidence of AI use in Shy Girl, a horror novel by Mia Ballard that generated significant buzz when previously self-published. The NYT brought their suspicions to the attention of the book’s publisher, Hachette Book Group. One day later, the publishing house decided to cancel the U.S. release of Shy Girl and discontinue the UK edition of the book, citing that the evidence provided violated their AI use policy. Shy Girl already sold thousands of copies in the UK, and thousands more had to be pulled from shelves after the story broke.
The decision to pull a book from publication has a range of major implications for a publisher like Hachette. Publishers invest a significant amount of time and money on acquiring, editing and marketing any book. Additionally, publishers often pay authors three advance installments throughout the publishing process. According to research by The Book Incubator, the average author advances from 2016 to 2021 were: $90,686 for memoirs; $65,908 for adult fiction; $50,718 for young adult; and $35,406 for middle grade. For a non-fiction book, New York Magazine reported hiring an outside fact-checker costs anywhere from $7,000 to $10,000 for a non-fiction book, depending on the length.
Shy Girl’s cancellation is widely considered the first time a major publishing house has pulled a novel specifically over AI use and could set the tone for how publishing companies deal with such issues moving forward, but not every publisher has the same idea of what level of AI use is acceptable.
Publishing companies face a wide range of AI risks, from copyright infringement to legal risk to reputation damage. To best mitigate these evolving risks, it is critical for publishing companies to understand the landscape of AI implications for the industry, establish the organization’s stance on acceptable or unacceptable use, and develop mature policies that can be communicated to staff, authors and customers.
Biggest AI-Related Risks in Book Publishing
The conversation around AI use and its risks in art remains ongoing and Hachette’s decision has brought a few of the key issues into the spotlight. Some of the biggest AI risks publishing houses face include:
Copyright Infringement from AI Use
If an author used generative AI at any point in the writing, editing or rewriting process and copyrighted works were used to train that AI tool, there is a chance copyrighted materials could end up in the final product. This can spell legal risks to the publisher for infringing on intellectual property.
The risk is compounded if the publisher does not find out about the use of any AI tools. While they could obtain proper permissions for the use of copyrighted materials, they have to know a rightsholder is even involved and discovering that requires a different level of scrutiny than one would usually dedicate to original works in the editing process.
If there is a copyright infringement issue, the publishing company could be sued, incurring defense costs and possible settlement or penalty payments. The resulting financial risk falls solely on the publisher. “A publisher has liability for the distribution of infringing materials," said Alan Heimlich, president of Heimlich Law. "While the author created the original work using AI, it is still up to the publisher to make sure the work complies with copyright laws before distributing the work.”
Lack of Copyright Protections
Another intellectual property risk is that AI-generated content cannot be copyrighted. “Under present U.S. copyright law, since no human authored the work, no one owns the work [and] no ownership will exist,” Heimlich said. “Therefore, these types of works do not qualify for the same protections provided by copyright. There exists a large void within the publishing industry because authors’ rights to their work remain unprotected.”
As a result, if someone copies the unprotected work, the author and publisher will have no grounds to fight it.
Reputation Damage
AI use also threatens to degrade a publishing company’s reputation in the eyes of readers and consumers. “Readers have already started suspecting undisclosed AI use from publishing houses, and they see it as a huge betrayal of their trust,” said Poppy Solomon, author and book editor. “They want real books by real people—work they can connect with emotionally. Now, they fear they are losing that. Not just because of AI, but the wider publishing landscape and what readers are dubbing ‘fast publishing’—the book version of fast fashion.”
Even accusations of AI use can harm the reputations of both the author and the publishing company. Such allegations can have devastating impact on an author’s audience, causing controversies on social media that can ruin an author’s career before it even starts. They can also raise questions about the publisher's work if it could have failed to detect AI use.
When a publishing house discovers AI was used, Solomon said it has a responsibility to investigate properly. “Ideally, the book should never get to the point of publication in the first place,” she said. “In Shy Girl’s case, many professionals along the way should have flagged the AI use—which readers found very clear—before it came to the witch hunt. Prevention is always better than a cure.”
How Publishing Companies Can Mitigate AI Risks
As both case law and professional ethical standards are still emerging, publishers should consider the following actions to mitigate AI-related risks:
1. Create AI Use Policies
While there is a burden on the author to either avoid using AI in the writing process or disclose it to potential publishers, it is ultimately up to the publishing company to detect AI and act accordingly if the use violates its policies. Since there is no single, agreed-upon law or guidance on AI use, individual organizations must create their own policies. For example, Hachette’s AI use policy states that all submissions must be original and authors must disclose to the publishing company if they use AI at all during the writing process.
That is a good start, but AI-use policies should also accomplish the following:
Defining AI-Related Terms
It is important to start with a clear understanding of AI terminology. “We have a no AI-generated content clause in our author contracts, but how we actually work with that in practice is more nuanced than a blanket statement saying that [authors] cannot [use AI],” said Jenn Grace, founder and publisher at Publish Your Purpose. “When an author comes to us and AI use has come up somewhere in their process, our first conversation is about understanding the distinction between AI-assisted and AI-generated. Those are meaningfully different, and confusing the two does a disservice to authors who are using AI responsibly and ethically.”
For example, “AI-assisted” might mean an author used a tool to help organize their thoughts, check grammar or brainstorm structure options. “AI-generated” means the prose itself came from an AI model with minimal human authorship. Defining these differences in a policy that is available to all staff, agents and authors gives all parties a common starting point for how the publishing company defines each AI-related term.
Providing Guidance on Acceptable AI Use
Once the AI-related terms are defined, publishers should ask themselves questions to determine what level of AI use they are willing to accept for a book to be considered for publication. Those questions may include: What if the book was only outlined with AI? What if the author just used AI to do research? What if it was only edited with AI? Is the only unacceptable case when the entire book has been AI-generated or is there nuance? “Many agents and publishers prefer AI has not been part of the process at all,” Solomon said.
Heimlich advised that both the company policy and the author agreement should identify AI uses that are permissible without prior written approval of the publisher, those that require prior written approval before the use and those that are not permissible at all.
2. Update Author Agreements
An author agreement is essentially a contract in which the author agrees to certain terms set by the publisher. It is critical to include AI-related indemnity and warranty clauses in these contracts and be sure to specify AI disclosure requirements.
“The author agreement should hold the author accountable for the final version of the work created through the utilization of AI and prevent the author from shifting accountability to the AI tool itself,” Heimlich said. “Additionally, publishers should incorporate indemnification language into the author agreement that requires the author to protect the publisher from any liability stemming from copyright infringement, plagiarism, providing false information, and utilizing content belonging to another person or entity without permission.”
Tim Billick, intellectual property attorney and partner at Practus, LLP, emphasized the importance of warranty clauses, recommending that a publishing contract require the authors to warrant that all of the work is original and does not infringe on the rights of others.
Critically, an author agreement must include provisions requiring the author to disclose all AI tools they utilized to create their research and outline, draft, revise and edit their work. This part should complement the guidance on what AI use the publishing company allows, and the author should disclose all AI use even if the publishing company does not allow it.
If a publishing company completely disallows AI use, Solomon said, “it has to be clear not just to staff but to agents and authors who are submitting work that AI books will not be accepted.”
3. Train Staff to Spot AI Red Flags
“AI is hard to ‘prove,’” Solomon said. “Most AI detectors do not work, [so] training staff to spot red flags would be a great help.” Unfortunately, it can be hard to train staff on spotting AI because AI was trained on human-written content. As a result, many of the “tells” can easily be flagged as AI-generated when they are not. However, there are some things editors or fact checkers can look out for.
“One of the funniest red flags you will see in a manuscript is someone leaving an AI’s response in,” Solomon said. “They copy and paste the AI-edited lines back into the book without deleting the AI’s agreement to the task.”
For example, Solomon said the scene might read like this:
Of course! Here is your dialogue, edited to be more realistic:
“Hey Jenny! Did you take the bins out?”
“I was too busy cleaning the kitchen!”
While that is a fairly easy red flag to spot, others are not so obvious. According to Solomon, some of the more common signs of AI-generated text are the excessive use of em-dashes and the passive voice. While the presence of one of these red flags does not guarantee that copy was AI-generated, it may raise questions that will require further investigation.
Hachette pulling Shy Girl could send a message throughout the publishing industry that AI-generated books are not acceptable, but it is not so straightforward. “It is hard to know what really goes on behind the scenes at a publishing house," Solomon said. "There are no rules when it comes to AI. All readers can do is keep pushing publishers to reject AI, while keeping in mind the impact an accusation can have. [AI use] is not fair to readers, authors or the greater publishing industry. A Big 5 publisher making a call like that will hopefully mean other publishers follow their lead.”