Publishing is fighting the wrong AI war

A $2 million book deal recently collapsed because nobody could answer what should be a fairly basic question in publishing: Who wrote the book?

Jerry Falade’s debut crime novel Call Me, I’ll Hide the Body had attracted huge interest. His agents ultimately withdrew the manuscript after saying they could no longer verify that it had been wholly written by their client. Falade strongly denies using AI and has challenged the allegations against him.

I am less interested here in whether Falade did or did not use AI. The part I find much more interesting is that publishing suddenly finds itself unable to answer a question that, until very recently, nobody thought required much infrastructure to answer.

An author wrote a manuscript. An agent represented it. A publisher acquired it. There was a chain of trust. Or maybe that is the better way to put it: there was always a chain of trust. We just never had to think very hard about proving what happened before the manuscript landed on someone's desk.

Generative AI is changing that. And I increasingly wonder whether publishing is responding by fighting the wrong battle.

Was AI used? That seems to be the question everyone wants answered.

Was the manuscript AI-generated? How much AI is too much AI? Should authors disclose it? Can we detect it? Should books get a "human authored" label?

The Wall Street Journal recently described an industry struggling with exactly these questions, from collapsed book deals to agents receiving what they believe are AI-generated submissions and publishers trying to figure out where assistance ends and authorship begins. AI detectors are being pulled into that discussion as well, despite the fairly obvious problem that getting this judgment wrong can have serious consequences for an author.

But what if "Was AI used?" is already becoming the wrong question?

Consider Allison Stanger, a professor and author who recently wrote in The Wall Street Journal that AI makes her a better writer. She does not ask the machine to replace her thinking. The argument, structure, memories, judgment and voice remain hers.

But she does use AI as an adversarial reader. She asks it to challenge her evidence, find weak transitions and argue the opposing case. She uses it for mechanical work. She even used two AI systems during the fact-checking process for her forthcoming book, while personally checking their suggestions against her original sources.

Is that an AI-written book? I don't think so. But AI clearly participated in making it. And drawing a nice clean line around that is going to get harder, not easier.

There is another reason I think we may be concentrating on the wrong thing. AI-generated books don't actually have to become better than human-written books to disrupt publishing. They just have to become abundant.

A study analyzing self-published genre-fiction books sold on Amazon showed that books with substantial AI-generated text generally captured fewer sales individually. But these books are reaching meaningful commercial scale. Their share of sales is increasing, and they are occupying more of the scarce top-ranking positions.

Even more telling: the number of books recording quarterly sales grew dramatically, while total quarterly revenue increased at a much slower pace. We are putting dramatically more books into a market where the money isn't growing at anything close to the same speed.

The researchers describe a market that can be reshaped through scale rather than quality. That, to me, is the more interesting part.

For hundreds of years, producing a book required meaningful amounts of time, skill and money. Those constraints certainly did not guarantee quality. We published plenty of terrible books before ChatGPT came along. But they created friction.

Generative AI removes a big part of that friction. And when we can create more and more content at almost no marginal cost, simply producing another piece of content becomes less interesting as a competitive advantage.

Something else has to become valuable.

This is where the discussion gets uncomfortable. What if readers actually like some of this stuff?

If someone buys an AI-generated thriller, enjoys it, finishes it and immediately wants another one, on what basis do we declare that experience illegitimate?

That doesn't make questions about copyright, rights, authorship or compensation disappear. Obviously not. But reader enjoyment isn't going to solve those questions for us either.

Publishing has spent decades telling itself that its purpose is to find and deliver content readers value. It would be a little strange if we suddenly decided that the process by which something was produced matters more than the reader's experience simply because a machine was involved.

And this isn't entirely hypothetical. Research on AI-generated literary excerpts has already shown that, under certain conditions, readers can rate AI output as highly as or even higher than work produced by expert human writers.

That does not mean ChatGPT can suddenly write the next War and Peace. It doesn't mean authors are obsolete. And we should be careful about extrapolating from a few short stories to the entire book market.

But it should make us cautious about assuming readers will automatically reject machine-generated work once it gets good enough. I don't think the future divides neatly into: Human books = good, AI books = bad. That would be wonderfully simple. I just don't think it is going to work that way.

Imagine an AI-assisted book that readers genuinely love. Great. That tells us something about its quality. It does not tell us who is responsible for it, whether the underlying rights are clean, whether factual claims have been checked, or whether someone else's work has been reproduced in ways it shouldn't have been.

Somebody still needs to stand behind all of that. And this is where the role of the publisher gets much more interesting.

I don't think publishers need a perfect system that tells them whether a manuscript is 7%, 28% or 63% AI-generated. Even if such a system became technically possible, I am not convinced the number itself would be particularly useful.

Take Allison Stanger again. If she has AI challenge her arguments, identify weaknesses and help fact-check the book, how much of the book is AI? Five percent? Twenty percent? Zero percent?

I have no idea. And more importantly, I am not sure I care.

The same problem appears quickly once we move beyond writing. What about translation, research, copyediting, brainstorming or fact-checking? At some point we risk building a very elaborate compliance process around a number that doesn't actually tell us whether the work is good, legal or trustworthy.

So maybe the better question is much simpler: Who takes responsibility for this work?

An author might have used AI extensively and still be prepared to say: these are my ideas, I made the decisions, I verified the claims, I have the necessary rights, I approve the finished work and I stand behind it. That seems much more meaningful to me than a box saying "no AI was used."

Fred Zimmerman makes an interesting argument here. Rather than building better and better AI lie detectors, he suggests thinking more in terms of chain of custody: understanding who worked on a version, which tools were used and who had permission to do what.

I like that because it turns a rather philosophical AI debate into something publishers can actually work with. We don't need a forensic reconstruction of what an author did at 3:17 p.m. on a rainy Tuesday while rewriting chapter seven. But publishers probably do need to know enough to understand what they are publishing and where the risks are.

For a long time, a publisher's logo quietly communicated something like: We selected this.

In an AI-saturated market, perhaps that statement has to mean more: We know where this came from. We understand how it was created. We believe the rights are clean. We checked what needed checking. And we are prepared to put our name behind it.

I am slightly hesitant about calling that a "trust layer" because it sounds suspiciously like something consultants would put on a PowerPoint slide. But the idea keeps sticking with me.

I don't want this to turn into another discussion about AI creating compliance work for publishers. There is a business opportunity here too.

Wiley reported $49 million in AI and data-analytics revenue in its latest fiscal year. McGraw Hill now serves more than 7.5 million active users across eight AI-enabled learning tools. Those businesses are obviously very different from trade publishing, and not every publisher can copy what they are doing. But the signal is interesting.

They are taking things publishers already own or understand well—such as content, rights, structured information, data and domain expertise—and making them more useful in a world where generic content can be generated cheaply.

Maybe publisher brands become part of that equation too.

Most readers probably couldn't name the publisher of the last five books they bought, and maybe AI won't change that. But if ten or twenty times as much content is competing for my attention, signals I trust become more valuable.

That signal might come from an author, a reviewer, a community or, perhaps, a publisher again. Not because the publisher controls access to the market—that ship has sailed. But because somebody credible chose the work, understands where it came from and is willing to stand behind it.

For years, the technology question in publishing has sounded roughly like this: What happens when authors no longer need publishers because technology allows them to create, produce and distribute books themselves?

AI certainly pushes us further in that direction. But there is another side to the story. What if the more content technology produces, the more valuable credible judgment becomes?

If we can generate 10,000 new thrillers tomorrow, producing thriller number 10,001 isn't particularly impressive. Knowing which three are actually worth reading might be.

That may be the paradox. AI could make several things publishers have historically done less valuable while making one of the oldest things they do more valuable: judgment.

What deserves attention? What can we stand behind? What deserves trust?

Which brings me back to the question I keep asking about AI and publishing: If content becomes effectively infinite, what is the scarce asset a publishing house actually owns?

It probably isn't its ability to create more of it. Maybe it is some combination of judgment, rights, audience relationships, reputation and trust.

I don't think that is bad news for publishers. It might actually make a great publisher more valuable than ever. But only if we stop defining our value by our ability to produce and distribute something that technology is making almost infinitely abundant.

The publisher of the AI age may be valuable not because it can produce content, but because it can make content trustworthy.

So maybe the question we should be asking is not: Was AI used?

Maybe it is: If producing more content is no longer scarce, what do we own that still is?