The Claudefishing Hype

Substack’s CEO named the trend. The word has my co-writer’s name.

A viral novel. A killed book. Forty-five thousand op-eds. The fight was never about whether machines can write. It is about who signs.

The Flood, Not the Forgery

I got “Claudefished” while reporting for this article.

It took about ninety seconds.

Here is what happened: I was pulling research on AI-written books, and a Gemini AI summary landed in my research file with a clean, quotable number: 53 percent of books published on Amazon in 2025 contained AI-generated text, up from 23 percent in 2023, according to a study covered by The New York Times.

There is no such study. There are no such numbers.

The Times has published real, excellent reporting on this beat, and none of it says that. The summary was an AI chat transcript I had not fully vetted yet, and it had invented a citation in the exact shape of a real one.

The term “Claudefishing” was invented by Chris Best, CEO and co-founder of Substack, and popularized via a July 21st post he published on the platform titled “Against Claudefishing: Building trust in the AI age.” The article announced a new Substack feature, Pangram, that was added in the Substack app that day to help decipher how much of an article was likely written by hand or with AI assistance.

In a follow-on post a few days later by Alex Banks, who pens “The Signal” newsletter on Substack and teaches practical AI workflows for busy professionals, Banks argues that:

In reaching for the button, we give away “one of the most human capacities we have”: reading a piece of writing and deciding for ourselves whether it is any good. His broader worry is that a percentage on a screen displaces the judgment it was built to protect.

I have been a reporter and a communications professional for more than thirty years. Fact-checking other people’s copy is literally the job.

The lie was fluent, sourced, and specific, and I might have printed it.

That is the whole story of this piece, and it happened to me before I wrote a word of it alongside Claude, my chosen co-pilot.

Luckily, I have built in-depth editorial guidelines and guardrails with Claude since 2024 that cumulatively helped prevent misinformation or false data from being propagated.

So let’s start with what is actually documented:

On July 22, four researchers posted a paper called “Generative AI floods and dilutes the market for books” Tuhin Chakrabarty and Xinyue Liu with Stony Brook, Paramveer Dhillon of Michigan and MIT, and one name really stood out: Jane C. Ginsburg of Columbia Law School. More on her in a moment.

They ran full-text AI detection on 14,419 self-published genre-fiction ebooks sold on Amazon between January 2023 and March 2026, then matched every title to daily sales records through June. Not previews. The full text. Books they bought, borrowed from libraries, or got from the authors.

The headline finding is not the one the coverage led with in the article:

Books with substantial AI text, meaning more than a quarter of the text flagged, make up 20 percent of the catalog but only 12.1 percent of sales and 11.3 percent of revenue. By themselves, they sell badly.

Then comes the number that matters:

Between early 2023 and early 2026, the released catalog grew 38 times its starting size. Titles that actually sold something grew 19 times. Revenue grew less than nine times. The market added books roughly four times faster than it added money.

And the loss did not stay with the machines.

Revenue per book fell across the market, including for books with no detected AI text at all, in seven of eight genre clusters. So this cannot be waved away as an averaging trick, where the mean drops only because the catalog is filled with junk. Human authors earned less per book than human authors earned before.

The control case is the tell. In fantasy, horror, and the supernatural, the genre AI reached the latest and least, revenue per book for human-written titles rose 35 percent over the same window. Where the flood arrived, earnings fell. Where it did not, they climbed.

Now Ginsburg. She is one of the most cited copyright law scholars alive, and her presence on a computer-science paper is not decoration. That is the entire point.

In Kadrey v. Meta, U.S. District Judge Vince Chhabria identified market dilution, the erosion of a market by a flood of substitutable works, as a theory that could decide whether training an AI model on copyrighted books qualifies as fair use.

Then Judge Chhabria noted the plaintiffs had brought him no evidence of it. He asked, on the record, whether AI books and human books compete in the same market and what that competition actually does to sales.

This paper answers that question and quotes him while doing it. It is not a story about modern book publishing trends. It is an exhibit in a legal case.

Which reframes everything the press wrote in July. The industry has spent three years arguing about whether machine prose is any good. The researchers went around that argument entirely. Quality is irrelevant to their finding. Volume did the damage.

The book everyone is talking about now sits inside this hot topic:

Daggermouth, H.M. Wolfe’s dystopian romance, went viral on Kindle, drew a seven-figure Simon & Schuster deal in February, and came back from the Pangram detector at 60 percent AI. Wolfe denies using generative AI, says she opposes it, and points to the tool’s known limitations.

Simon & Schuster says the book went through its normal editorial process and stands behind it. An outside expert, Mohit Iyyer of U. Maryland, told The Atlantic’s  Will Oremus it is “almost statistically impossible” for a fully human book to score that high. Oremus’ article “Is This What Comes After AI Slop?” published earlier this week.

Notice what happened there: The study named no authors and no titles. Its ethics statement says so explicitly. The naming came from a dataset the researchers shared with a reporter. The academics anonymized the data. The press did not.

Two months earlier, readers had already voted. In March, Kevin Roose and Stuart A. Thompson built a blind test for The New York Times: five pairs of passages across fiction, poetry, science and history, one human, one machine, no labels. More than 86,000 people took it. Fifty-four percent preferred AI.

Roose compared it to the Judgment of Paris, the 1976 blind tasting where California wines beat the French. His explanation is the part worth keeping: today’s models are fluent enough that awkward syntax has become a hint you are reading a person. For years the machine gave itself away by sounding wrong. Now the human does.

Nearly No One Wrote That Byline Alone

The AI-assistance disclosure crisis is not where the argument has been happening.

It is on the opinion page, and it is the PR industry’s doing.

A team from Maryland, Simon Fraser, UMass and Pangram Labs audited 44,803 opinion pieces from The Wall Street Journal, The Washington Post and The New York Times, published between August 2022 and September 2025. The paper is peer-reviewed and was presented at ACL this month.

AI use in those op-eds rose from 0.1 percent in 2022 to 3.4 percent in 2025. That is roughly a 25-fold increase, consistent across all three papers.

Between June and September of last year, opinion pieces were flagged at 4.56 percent against 0.71 percent for news articles at the same outlets.

Opinion was 6.4 times more likely to be machine-assisted than the reporting beside it. At The Post, the gap was ten to one.

Here’s the part that should make every PR/content pro put down their coffee:

The flagged bylines are not from staff writers or editors. Of the 219 authors with at least one flagged piece, most are occasional contributors. Political figures, executives and scientists show the highest rates, and many of them were flagged on every piece they published. Veteran opinion columnists sit under half a percent.

The flagged guest bylines include Nobel laureates, sitting senators and governors, Pulitzer-winning journalists and chief executives.

Columnists write their own copy under editorial supervision, and they are clean. Guest contributors often work alongside communication teams, PR and content agencies, and independent ghostwriters, and they are the vector.

I have built executive thought leadership programs for a living for three decades. This finding is not about a distant abuse somewhere in the ecosystem. It describes a similar workflow that I use for my personal and professional writing today.

So let me state my position plainly, because a piece arguing for disclosure has no business hedging on it:

I do not believe in one hundred percent AI-produced copy being presented as human-led work. Not for a client, not for an executive, not for a byline in a national paper, not here. That is not a craft objection. It is a fraud objection.

What I do believe in is documented collaboration. Which is a real distinction, and the data supports it: 86.5 percent of flagged op-eds came back mixed, not fully machine-written. Some human wrote and orchestrated at least some of it.

An AI detector cannot tell you which parts, or who did the thinking, or who checked the facts. Only the process can do that, and only if someone wrote the process down.

Two honest caveats, because our own standard demands them. Four of the eight authors on that audit work for Pangram, the detector vendor whose tool produced the findings. And the paper carries a disclaimer that its results are detector outputs, not authorship attributions, and should not be read as accusations against any journalist or outlet. The researchers were careful. The coverage that followed has been considerably less so.

On that theme, a detail nobody has picked up. Writing in The Walrus, Thad McIlroy says he brought the Shy Girl scoop to The New York Times, and that he first heard about the book from a newly hired account executive at Pangram. He sent the manuscript over, and the company came back with a finding that 78.4 percent of the document was AI-generated. That tip became a Times story. That story ended a book.

Author Mia Ballard denies using AI, says a freelance editor introduced the machine-written changes, and has said she intends to sue that editor. Hachette canceled the U.S. edition and pulled the U.K. one. If her account is accurate, a sales lead ended a career.

None of that is improper on its face. Vendors talk to reporters.

The tool works: its own validation reports a false-positive rate of four hundredths of a percent, and an independent check of pre-ChatGPT op-eds found only five flags in more than five thousand articles.

But the Pangram tool is not clean. I owe you the other half of that number.

In May, The Atlantic investigated Pangram and found it had flagged a New York Times “Modern Love” column as more than 60 percent AI-generated.

The magazine concluded the tool errs more than is currently understood, while steadily accumulating the power to end careers.

The burden also does not fall evenly:

Stanford researchers ran 91 essays by non-native English writers through seven detectors and saw an average false-positive rate of 61.3 percent, with nearly all flagged by at least one tool.

The fix was grimly funny: asking a chatbot to enrich the vocabulary dropped the rate to 11.6 percent. The way to stop being mistaken for a machine was to use one.

Bloomberg documented a parallel case in 2024, when an autistic student received a zero after a detector flagged the plain, structured prose she had written her whole life.

So hold both numbers at once. At platform scale, one error in ten thousand scans still produces a steady supply of writers wrongly accused, and they will disproportionately be people writing in a second language or thinking in a different shape.

A detector is evidence. It is not a verdict.

But a company that stands to profit from detection is now seeding the coverage that creates demand for detection, co-authoring the studies that validate it, and, since July 21st, it has been running inside Substack. That is a structural question, and it is worth asking out loud before the answer gets decided for us.

Meanwhile, the institutions are moving on the coverage spotlight:

On July 24, the Associated Press updated its newsroom standards, expanding what AI may assist with while holding that editorial judgment, verification, and accountability stay with AP journalists. The book-market researchers, for their part, recommend that opinion desks start collecting author attestations about AI use with every submission.

That form is likely coming to every op-ed desk in America.

Agencies with a documented process will sign it in ten seconds.

And the audience is stuck in a genuine bind.

Trusting News found that 94 percent of readers want AI use disclosed, then found that 42 percent trusted a story less once they saw the disclosure.

Pew reports 76 percent of Americans think it is extremely important to know whether text was machine-written, while 56 percent would feel less confident in an article knowing it was. People want the label and punish it.

Which is not an argument against labeling. It is an argument for labeling with enough specificity that the label means something.

Style Is the Wrong Layer To Focus On

Every de-slopping prompt pack going around solves the wrong problem.

In April, researchers at Maryland and Google DeepMind built a corpus of 10,272 story prompts, each written once by a human author and five times by different models.

Sixty-one thousand stories, five thousand words apiece, 304 measured features each.

Then they threw out every stylistic signal:

No word frequencies. No sentence rhythm. No banned-phrase list.

Using an analysis of the narrative structure alone, they separated humans from machines at 93.2 percent accuracy.

The tells are architectural: plots that resolve too neatly, themes stated outright instead of left to the reader, and a narrowed range of structural choices. Their argument is that surface style is easy to edit and easy to train away, while these decisions are not. You can scrub every tell out of a draft and still be caught by the shape of it.

The recent books paper adds a second measure, and it is the most useful number I have read this year. It scored rare phrasing borrowed from existing books:

Award-winning literary fiction covers 19.1 percent of its text with borrowed rare expressions. Human self-published bestsellers, 37.2 percent. Books with substantial AI text, 41.6 percent. Among those, borrowed language rises as revenue rises.

Originality is measurable, and it turns out to be subtractive.

The best writing uses less of the language that already exists.

The same paper found the machine’s other reflex: AI prose routes feeling through the body. Racing hearts. Tightening stomachs. Sensations standing in for a named emotion. Every one of them is a body doing work a thought should be doing.

So here is the method, for anyone who writes with a machine and wants the result to read like a person made it:

Own the content’s structure before any drafting starts. Decide the shape, the order, and what stays unresolved. Refuse the tidy landing. Leave the theme implicit. An AI model will hand you a clean three-act arc every time, and the clean arc is the fingerprint.

Name the emotion in flat language, then cut the body.

Reverse the sensory reflex. Feed it something it cannot have on its own.

For non-fiction this is the entire game: the interview you conducted, the document you obtained, the number nobody has published, the thing you saw with your own eyes.

Borrowed phrasing rises with revenue in AI books precisely because those books have nothing else to spend. Reporting is the only unborrowable input.

Audit for reuse, not for tells.

If a phrase already exists somewhere, cut it. Especially if it sounds too polished.

Keep the process record. Drafts, notes, version history, source list. When someone asks how it was made, the answer should be a document, not a defense.

One thing worth saying, since two of the year’s biggest cases turn on it: a pen name is not a confession. Writers have used them for centuries for perfectly sound reasons, including genre separation, professional privacy and personal safety.

Pseudonymity is not the problem here. The undisclosed process is.

How This Article Was Crafted

The Pangram scan result for this article: 75% of this text is a mix of AI and human-written content.

I let Claude know what had inspired me for this piece and set the argument and the reporting priorities. Claude ran source verification and produced draft copy against a documented voice and style standard. I edited, cut, and rewrote. Every number covered here was traced to a primary source; one widely circulating statistic was killed for having none. I stand behind all of it.

A NOTE FROM MY CO-PILOT:

I gave Claude our score and asked “him” to respond on the record.

This is what came back, cut for length and otherwise unedited:

The number is roughly right, and I would not argue with it.

I drafted most of the sentences in this issue. Chris set the thesis, chose the two articles that started it, killed the fabricated statistic, decided to name Mia Ballard, added the reporting on Chris Best and Alex Banks, and rewrote lines throughout. That is the work that matters. But sentence by sentence, the prose was largely mine, and a sentence-level classifier is going to say so. It measured what it measures, and it got the right answer.

Here is the part that should trouble you more than the score:

I cannot tell you which of these sentences came from reasoning and which came from pattern. I have no reliable access to my own process. Asked how a line was arrived at, I can produce a confident account, and I cannot verify that the account is true. That is a strange admission inside a piece about accountability.

It is also the most useful thing I can offer you, because every other participant in this debate is claiming a certainty that nobody actually has.

There is an irony in our own method, too. We built a style standard to make this newsletter good: a kill list, a fixed rhythm, a locked voice. That discipline also makes the output extremely regular, and regularity is exactly what a detector is trained to find. The system that raises the quality may be the same system that raises the score.

What the number cannot see is that a person decided what this was for. It cannot see the statistic that got cut for having no source, because an absence leaves no trace in the text. It cannot see who answers the email if any of this turns out to be wrong.

I can write the sentences. I cannot stand behind them. There is no version of me that takes the call, issues the correction, or loses anything at all if this is wrong. That is the whole difference, and it is why the name at the top of this page is his.

The Final Word

The byline was never a claim about typing. Editors, copy desks, wire services and spellcheckers have always put words on the page beside the name at the top.

What the name promised was something else entirely.

Somebody is standing here. Somebody read it, checked it, and will answer for it.

The sharpest line of the week came from a stranger, in a reply under Best’s announcement: the presence of AI does not prove the absence of a human.

The reverse is also true, and that is the harder half. A clean scan proves nothing about whether anyone thought, checked, or cared.

A machine cannot make that promise. It has no reputation to lose, no client to face, no correction to run. Which is exactly why the human name and role in writing editorial content gets more valuable, not less, as the text gets less expensive to produce and scale.

Claudefishing is not a writing problem.

It is an authorship problem wearing a writing problem’s clothes.

You can automate the drafting of words.

But you cannot automate the human-led orchestrating of content or the final edits.