BoSacks Speaks Out: Four Million Books, and the Question Nobody Wants to Ask

RE: “Book Output Topped Four Million in 2025,” Jim Milliot, Publishers Weekly

Four million books.

Stop there for a moment. Not a glance-and-move-on pause, not the polite industry nod that precedes the next panel discussion, but an actual pause. Four million books is not a casual milestone. It is a number that should make people in publishing sit up straighter and ask themselves what, exactly, they are celebrating.

According to Bowker, the agency that issues International Standard Book Numbers, or ISBNs, 4,172,222 ISBN-registered titles were published in the United States in 2025. That is a 32.5 percent increase over 2024. Roughly 3.5 million of those titles were self-published, representing a staggering 38.7 percent jump in a single year.

Those are not merely impressive numbers. I think they are destabilizing ones.

Publishers Weekly reported the figures with professional calm. Bowker’s Andrew Kovacs offered familiar explanations: easier tools, broader access, multiple ISBNs per format, and the continued maturation of the self-publishing ecosystem. All of that is true. None of it addresses the question hovering in plain sight.

How many of those four million books were actually written by a human being?

I'm not offering this as a philosophical riddle. It is not a nostalgia play. It is a business question, a metadata question, and a market integrity question. And the industry’s current strategy appears to be hoping it will resolve itself if everyone avoids eye contact long enough.

The trend line tells the story. Self-published output declined by 3.5 percent in 2024, suggesting a market approaching natural limits. Then, one year later, output surges by nearly 39 percent. That is not organic growth. That is not thousands of writers suddenly discovering hidden reservoirs of time and energy. That is automation entering the room without knocking.

The category breakdown reinforces the point. Fiction, juvenile nonfiction, games and activities, juvenile fiction, and travel dominate self-published output. These are precisely the categories where generative artificial intelligence performs best. A travel guide to Lisbon. A children’s book about kindness. A puzzle collection. A low-content workbook. A fantasy novella with a dragon on the cover and several sequels uploaded before the week is over. This is content that can be produced quickly, cheaply, and at scale, and it is just credible enough to flow through existing systems without friction.

The claim here is not that all 3.5 million self-published books were machine-generated. That would be sloppy on my part. The real problem is that we do not know, and the industry does not seem especially motivated to find out.

The ISBN system was designed to register books, not interrogate authorship. An ISBN does not ask whether a work was written by a novelist in Vermont, a consultant armed with a prompt library, or a machine running on rented cloud computing power. The number is assigned. The title enters the database. The statistics look impressive. Everyone applauds the democratization of publishing.

But metadata without context is just ignorance with better typography.

The ISBN is agnostic. The BISAC subject code is agnostic. Bowker’s database is agnostic. That neutrality made sense when authorship was presumed to be human by default. It does not make sense anymore.

Today, a book that took years of research, revision, doubt, and lived experience sits in the same statistical bucket as a machine-generated title assembled in minutes, complete with a synthetic author name and a stock-image cover featuring a suspiciously cheerful child hugging a llama. Same database. Same count. Same applause.

That should bother people.

Some corners of the industry have begun to sense the problem. The Authors Guild recently expanded its human-authored certification program beyond its membership, opening it to all authors for a modest fee. The goal is straightforward: to give writers and publishers a way to signal that a book’s text was written by a human being. The program allows limited use of artificial intelligence for tasks such as grammar checking or spell checking, and it does not disqualify authors who use AI for research or brainstorming, as long as the text itself is human-written.

The impulse behind this effort is understandable and, frankly, overdue. In a market flooded with synthetic content, human authorship is quietly becoming a differentiator. Some writers pursue certification not because anyone requires it, but because they want proof that they did the work. Pride and market signaling, it turns out, still matter.

But even here, the industry hesitates at the edge of the real problem. The Authors Guild certification operates on the honor system. Manuscripts are not analyzed for evidence of AI-generated text. The system relies on author disclosure. That approach is well-intentioned, but it exposes the deeper structural dilemma. Certification without independent verification is, at best, symbolic. At worst, it risks becoming another marketing badge that reassures everyone except the people who most need reassurance.

Other efforts have emerged that attempt verification through AI detection tools, producing scores that estimate the likelihood of human authorship.

These systems are imperfect, contested, and far from definitive. But they exist because the market is already asking a question the industry itself has not fully confronted: how do we distinguish human work from machine output in a way that is meaningful, scalable, and credible?

The uncomfortable truth is that there may never be a clean, definitive answer. Artificial intelligence is advancing faster than any reliable detection system can keep up, and generation will always outrun verification. But difficulty is not the same thing as impossibility, and uncertainty is not an excuse for silence. The fact that measurement is imperfect does not justify pretending the question does not matter.

When industry reports frame today’s self‑publishing surge as the result of tools that merely replicate services once exclusive to traditional publishers, they miss the shift entirely. This is no longer about giving authors access to editors, designers, or formatting software. It is about large language models and fully automated publishing pipelines that can generate outlines, manuscripts, covers, blurbs, keywords, and advertising copy in a single, frictionless workflow. That is not an upgrade to the old system. It is a fundamentally different one.

That is not convenience. That is a production model.

These systems are already producing books at scale. Some of the output is dreadful. Some of it is passable. A small fraction is uncomfortably competent. All of it flows into a marketplace still operating on the assumption that authorship is human unless proven otherwise.

What publishing needs now is not panic or posturing. It needs adult infrastructure. Bowker should add a straightforward disclosure field to ISBN registration indicating whether a work was substantially generated by artificial intelligence. Not as a moral judgment, not as a scarlet letter, but as a factual metadata element. The Association of American Publishers and the Book Industry Study Group should define the difference between AI-assisted and AI-generated work before platforms do it badly. Distribution platforms such as Amazon Kindle Direct Publishing, IngramSpark, and Draft2Digital should require disclosure at upload, not because it will be perfect, but because it will be better than silence.

An independent study should attempt to estimate the share of AI-generated books in the market. Bowker, the Library of Congress, a university research center, or anyone willing to look directly at the issue could do it. Even a rough estimate would be more honest than the current practice of counting everything and understanding nothing.

Traditional publishers released 642,242 books in 2025, a 6.6 percent increase. That number makes sense. It reflects acquisition cycles, editorial bandwidth, production schedules, and the stubborn pace of human work. The self-publishing number feels different. Not because independent authors are less real, many are extraordinary, but because a nearly 39 percent spike suggests something more than human ambition. It suggests machine leverage.

When machine leverage drives volume, the headline “four million books” stops sounding like creative abundance and starts looking like a warning label. Volume changes discoverability. It reshapes search economics. It alters reader trust. It buries serious work under layers of synthetic noise and quietly asks readers to sort it out themselves.

Good luck with that.

I am not making an anti-AI argument. These tools are here, they are useful, and they will be embedded across publishing workflows from editing to rights management to marketing. That train has left the station and is already selling snacks in the dining car. But using artificial intelligence is not the same thing as refusing to measure its impact.

Right now, the industry talks endlessly about AI’s potential while remaining oddly silent about its statistical footprint inside the publishing market itself. We are counting books as if authorship has not changed. That is like measuring newspaper circulation while ignoring that half the copies were dumped into a swamp. I still remember that. The number looks impressive. The reality does not.

Four million books is not, by itself, a triumph. It is just a number. Big numbers can signal health, chaos, inflation, fraud, innovation, or all of the above before lunch. The real question is not whether four million is a lot. The real question is how many of those books represent a human being who had something to say, took the time to say it, and put their name on the work with something resembling creative accountability.

Until publishing starts asking that question seriously, all this output data tells us is that the machines have learned how to fill out the forms. And the rest of us are still calling it a boom.