The Accusation Is the Product

TL;DR: Dan Holloway set two seven-figure deals side by side for the Alliance of Independent Authors. H. M. Wolfe’s Daggermouth, “the highest-profile indie bestseller of the past year,” took a study finding “that the book may be 60 percent AI-generated,” kept its Simon & Schuster deal, and debuted at number one on the New York Times list. Jerry Falade’s Call Me, I’ll Hide the Body came out of a reported $2.4 million deal after his own agency, which never ran a detector, cut him loose. The one with a score against her is fine. The one nobody tested is not. The paper underneath both sums up its 14,419-book corpus in a line: “None of these books disclose whether or not they contain AI-produced content.” That gap covers your product page too.

I went in looking for which one did it. Wrong thing to ask. Both deny it, both denials are unfalsifiable, and they split anyway. What split them was sitting in the sales data.

Two writers, one allegation, two endings

Daggermouth scored 60 percent on the AI detector Pangram. Chakrabarty’s team at Stony Brook then ran a second check for rare expressions, strings that appear “in at most five Google Books volumes” and nowhere in a 4.7-trillion-token sweep of the internet. The book was thick with them.

Wolfe denied it through her lawyer. Simon & Schuster stood behind the book. Mohit Iyyer, who studies detectors at Maryland and had no hand in it, told the Atlantic it is “almost statistically impossible” for an entirely human-written book to register at 60 percent. The hardcover shipped July 28, the day the study became national news. It debuted at number one.

Falade never got a test. His agent Marc Gerald told the Bookseller an editor raised concerns midway through submissions, and that Europa Content declined to run a detector because Falade was “clear and credible.” Then a meeting on 29 July at which, per Gerald, “aspects of [Falade’s] story changed.” Sandy Hodgman emailed international publishers: “we are no longer able to authenticate how the manuscript fully evolved from origin to completion.” The agency cut ties.

Fourteen houses had been bidding. Futurism reports the $2.4 million came from Minotaur, which worried about securing copyright in a manuscript that might carry machine-written passages. Falade on the clock: “rumours began circulating and, within six hours, I had been dropped by my agency, deals were paused, and press releases were being issued.” And on the pattern: “There seems to be a troubling assumption that when a Black writer produces work that attracts significant attention, the work could not possibly be their own.”

What’s on paper points harder at Wolfe, and it is not close. A detector output, a second test on a different signal, a specialist with no stake saying the number should not be possible.

The one with the file kept the deal. The one with nothing lost it in an afternoon.

What separated them is where each book already sat. Daggermouth was in stores, earning, sequel scheduled, with readers who read every attack as proof the book matters. Falade had a pub date, a stack of promises, and not one paying customer. An unproven claim gets whichever answer is cheapest, and how far along the book is decides which answer.

What the study set out to do

Holloway’s follow-up two days later is the corrective the rest of the coverage skipped, and he does it by reading the title out loud. The paper is called “Generative AI floods and dilutes the market for books.” Its aim, he writes, “is not to name and shame indie authors (or others) for using AI but to describe a landscape in which the use of AI might diminish the ability of those who don’t use the technology to sell.”

The paper bears him out. They pushed the full text of 14,419 self-published genre-fiction books on Amazon, 2023 to 2026, through a detector and matched each to daily sales. Books with substantial detected AI text take a large share of the catalog, a smaller share of the money, and a growing share of the top ranks. Supply is what does it: “the number of books with observed sales in a quarter grew 19.2-fold, while quarterly revenue grew only 8.9-fold.” More books chasing the same money.

Then the abstract states the job. “Our results bear directly on the market-effect question at the center of the fair use defense to copyright infringement.”

It is a copyright brief. The coverage turned it into a suspect list.

The listeners were not told, and they preferred it

Here is the finding nobody in the panic wants, and I will name its funding in the same breath. Edison Research at SSRS ran it. Spoken, an AI audiobook company, paid for it. In the blinded test, more than a thousand US fiction listeners heard the same thriller from a professional narrator or a one-click AI multi-cast. On character-driven scenes listeners liked the AI better, 61 to 53, and rated it more engaging, 58 to 49. Sixty-one percent of the people who got the AI thought it was a person.

Discount all of it for who wrote the check. Something survives. The human narrator won the one-voice stretches that explain instead of talk, and took purchase intent 49 to 46. Three points is not a moat.

Dervla McTiernan names the same collapse in CrimeReads. Nine months ago she could still tell AI images and video from the real thing. “That time has now passed.” Then the line that belongs over every acquisitions desk: “accepted truth becomes less about what can be proven and more about who is believed, how often a claim is repeated and whose version of events reaches us first.”

She wrote that about deepfakes inside a crime novel. By accident she wrote down how publishing decides who did it.

Write the disclosure before anybody asks for it

Write your disclosure before anybody asks for it. Five moves, none of which costs money.

Date it and own the page. A dated disclosure on your own domain is a record. A tweet is not. Ours has been up since Pangram raised its round, the only reason I get to give this advice with a straight face.

Be specific enough to be boring. “AI-assisted” reads like a man confessing to the smallest thing he can get away with. Name the tool and the job. Research triage, transcript cleanup, alt text, nothing at all. A boring detail kills a story.

Keep the paper trail. Hodgman said the problem was authenticating how the manuscript evolved. A finished file cannot answer that. Dated drafts can, and they cost you a folder of timestamped exports.

Put it on the product listing. A policy page nobody opens will not reach the person deciding whether to refund.

Do not reach for a detector to settle it. Europa declined to run one and dropped Falade anyway. Pangram ran on Wolfe and she is number one. A false positive you generated yourself is a piece of paper that now exists.

The part I would argue with

Two things, and the first is a line Holloway wrote himself.

In the piece setting the two deals side by side, he writes that Falade’s deal fell “after the publishers felt they could no longer substantiate that it hadn’t been AI-assisted.” Look at whose name is on the emails. The notice went out over his own agents’ letterhead, and they cut him loose inside a working day. Minotaur was worrying about copyright, per Futurism, so the imprint was hardly passive. Still, the people who moved first were the ones he was paying. If publishers were the whole danger, a contract would be the shield. No clause protects you from the people you hired losing their nerve on a Wednesday.

Second, he says the study’s aim “is not to name and shame.” Publish a tool and you stop owning what you meant by it. The market-effect prong of fair use needs you to name each title: which book, which detector, what percentage. That builds a naming machine whether you meant one or not. One reporter worked with the researchers, pulled a single name out, and the name carried a magazine feature.

I also have to move where I started: indies would eat the damage first, since they keep no lawyers and no publicist. These two cases say otherwise. Wolfe is the indie, with more against her on paper, and she is fine. Falade had trade money from Macmillan and he is not. A book people already pay for turned out to be worth more than any big name behind you.

Frequently asked questions

Does disclosing AI use invite the accusation rather than preventing it?

It removes the story. Accusing somebody is news because it promises a reveal, and a reveal needs something hidden. If your process is posted, dated, and specific, the most anyone can produce is a screenshot of a page you wrote. Falade’s agents never claimed proof. They said they could no longer authenticate how the manuscript evolved, and that gap only opens where nobody documented anything.

Isn’t the honest answer just to not use any of it?

Fair enough, and Wolfe’s flat denial did not stop anybody accusing her anyway. Detectors return false positives on human prose, and you have to disprove an output you cannot inspect. So if you use nothing, say so on a dated page, and keep your drafts anyway.

If detectors produce false positives, why keep records at all?

Because the counterparty is almost never asking whether a detector flagged you. Minotaur’s reported concern was copyright. Hodgman’s was authenticating how the manuscript evolved. Both are origin questions, both answered by process artifacts, neither by a detector score. Old drafts, dated files, research notes, old email: cheap to keep, impossible to rebuild after somebody asks.