“This AI moment—in the industry, in the world—is revealing a lot about what we believe to be important,” observes Rakuten Kobo CEO Michael Tamblyn, “and about why we all do the jobs we do.”

By Michael Tamblyn, President and CEO at Rakuten Kobo

Iwant to tell you about an argument that happened recently in my office—a messy, multi-sided, very vigorous debate.

We were having our regular meeting, where we get together to solve the problems that need more than one or two people to solve. The people at the table are responsible for pretty much every facet of our work at Kobo—finance, marketing, publishers and authors, products, technology, sales, operations, and legal. Some of us come from the book business, some from the tech side, different backgrounds, different countries. And we like each other. We are not only colleagues, but friends.

So, what was the subject that got a group of friends and colleagues who don’t usually argue to start arguing? We were discussing whether we should sell a book in our bookstore if it was made with AI.

At Kobo, we have millions of titles in our catalog, and we receive thousands of new books every day. And the number of books that look like they are being made by AI (or with a lot of help from AI) is growing. We all watched with great interest as the controversy unfolded over the cancellation of Mia Ballard’s Shy Girl at Hachette.

Against this backdrop, the various teams at Kobo were looking to establish a policy: which books should we let in? Which should we keep out? And how do we decide?

Of course, we make these kinds of decisions all the time. We don’t allow hate speech, for example, or pornography, and we try to catch false or misleading medical information. And we have a set of rules that we publish so that everyone knows what we will sell and what we won’t. That’s part of being an online bookseller.

So why was this discussion about AI generated works different? Why did it generate such strong feelings? Because this AI moment—in the industry, in the world—is revealing a lot about what we believe to be important, and about why we all do the jobs we do.

Plutonium, or Salt?

In retrospect, our debate over AI at Kobo was really three debates in one, each with its own challenges: First, for a book, how much AI is too much? Second, if we’re going to make rules about AI use in books, how do we enforce those rules? And finally, in a world with AI in it, what kind of bookstore do we want to be?

For reference, we receive thousands of ebooks every day from independent authors through our self-publishing service, Kobo Writing Life. This is also where we see a lot of AI-generated books being submitted. Last year, we rejected 45% of the self-published books that were submitted—literally hundreds of thousands of books—and most of those rejections were because, best we could tell, the books were generated with AI.

They were also terrible. If a customer bought one, we concluded, they would surely be disappointed and want their money back and be angry at us for selling it to them. In those cases, the decision is easy: a poor-quality book clearly created with AI should not be allowed in the Kobo store.

At the other end of the spectrum, the decisions are also easy: A good book written and/or edited by a human, whether self-published or from a traditional publisher, should go in the store so long as it doesn’t break any of our other rules. That gives us two ends of a spectrum—we don’t want to sell very bad, obviously AI-generated books. Quality, human-authored books, we absolutely do want to sell. The problem is everything in the very sizable middle.

The fact is, many authors today are using AI tools. Nonfiction authors, for example, are using AI as an assistant to organize notes and to collect and summarize research before they start writing. We also know that many authors are using AI in their writing process. For example, an author struggling with a certain passage might ask an AI tool: “Write me five different versions of these paragraphs, keeping my tone and style, and the same number of words.” The author may not use any of the AI generated rewrites, but by reading them, maybe she finds inspiration for the version that ultimately writes herself.

But what if she does use one of them? Or, what if (as is alleged with the Hachette title, Shy Girl) a substantial amount of the book was written—or edited, or rewritten—with the assistance of AI? We also know that editors, especially at small publishers that lack resources, are using AI tools to check spelling, grammar, and punctuation. How should we factor that into our thinking?

What we realized in our debate is that we were arguing over a specific question: For books, is AI like plutonium? Or is it like salt? Plutonium is deadly in the smallest quantities. Salt can kill you if you use too much of it, but in small quantities, it makes things better. Another way of saying it: should AI be treated like plagiarism, where the presence of it, no matter how small, means a book should be disqualified? Or is AI a tool that can be used in moderation?

For many working authors today, AI is plutonium. After all, we know that many of the most popular AI services today were trained using pirated books, and authors and publishers are angry about that. Even when we talk about using AI to help with book recommendations, search, or discovery, the question from many in the industry remains: “why would you use tools from companies that stole our books?”

But we also know that more authors and publishers are, in fact, using AI tools. So, if we say that we will only sell books with 0% AI, we will likely end up disqualifying many books that deserve to be read. Somewhere between low-quality AI slop and Cervantes, Borges, Balzac, Sarah J. Maas, and Zadie Smith, there is a line. But where to draw it?

Boundaries

The second part of our debate was no less thorny. Once we decide to draw a line and create a rule, how can we enforce that rule with any accuracy?

After all, any rule we might make regarding AI use can only be enforced if we can detect AI use. Not long ago, detecting AI wasn’t very difficult. AI generated content didn’t look like a human wrote it—it was weird, with strange grammar and sentence structure, the text equivalent of an AI photo with too many fingers.

But as the models continue to rapidly improve, it is getting harder and harder to detect AI generated text, and it could soon be impossible. Each time a new AI version is released, the quality of text generation gets significantly better, and the task of detecting AI use becomes more challenging.

Yes, there is a growing market for software that purports to tell us whether a book contains AI. But so far, the software Kobo has evaluated regularly mistakes text written by humans as being written by AI, or vice-versa. For us, the risk of relying on software to make a determination that might damage an author’s career by falsely labeling their book as AI is still too high. Further complicating any rule we might make, none of the tools we have evaluated are very good at telling us what percentage of a text was created with AI.

One simple strategy: we could ask authors and publishers to certify that their books are created without AI. But while we can be reasonably confident that the authors who don’t like the technology—and most of them really don’t—will be honest, our experience suggests that those who want to shovel thousands of machine-made books into our catalog won’t be.

Meanwhile, another interesting piece of research also caught our attention. A recent article in a popular linguistics journal found that people who work frequently with AI text creation are far better than AI detection software at evaluating whether something was made with AI. This suggests an interesting conundrum: if we keep editors away from generative AI, that only makes them less likely to be able to detect when a manuscript that was written with AI, and makes it more likely that AI-generated work will creep into the publishing pipeline.

Choices

Whether one is a small independent bookstore or an online retailer with millions of titles, we all make choices. We screen some books out because we don’t want them in our store. Thus, ultimately, at the heart of the third part of our debate over AI is a question of values: In an AI world, what kind of bookstore do we want to be?

The core of this debate resides between two viewpoints, which I can summarize as “Readers First” vs. “Book Community First.”

The position of “Readers First” is that, if our job is to find readers their next book to read, why should we prevent them from reading good books they might like just because they have used AI in their creation? “Readers First” says if the book is good and readers want it, we should do our job and offer the book.

On the other hand, the position of “Book Community First” says that we must not only support readers, but the entire ecosystem of authors, publishers, and editors. If we don’t make any rules, we risk letting AI created works drown out new, innovative voices under a pile of machine-generated text. That not only costs people their livelihoods, but it also makes it harder for new, original ideas and stories to be written.

In our debate, we ended up leaning more on the side of “Book Community First.” We need authors and publishers to be able to make a living. That’s one thing we could all agree on: we only have a book industry because people value words on a page, and because people are willing to spend time and money for the pleasure of reading them.

That conclusion doesn’t make coming up with new rules about the complex, and evolving use of AI in the book world any easier. But it is certainly clarifying: as booksellers, we must do our part to take care of authors and their works.

This piece was adapted from from Michael Tamblyn’s keynote at the Frankfurter Buchmesse’s CONTEC Buenos Aires edition, delivered on April 21, 2026 during the 50th Buenos Aires International Book Fair.