Claudefishing
Where the line actually falls between authoring with AI and letting it author for you.
Cover image: Thomas Edison with a dictating machine, 1907. Photographer unknown. Public domain. Via Wikimedia Commons
YEARS AGO I sat down and explained Shariah to a room of serving police officers in Houston. What the word means. Where it comes from. What it governs, what it does not require, and why almost everything said about it in public is wrong.
I wrote that text myself, every sentence of it. There was no assistance to be had.
Later I took that text and turned it toward politics, because that is where the argument had moved. I reorganised it, reframed it, brought in what had happened since. It ran on Fireline Press.
In July, Substack gave every reader an AI detector: a button on any post that returns a percentage — how much of it a human wrote.
I ran it on that piece. The machine read my teaching text and told me it was almost entirely machine-written.
I want to be careful about what I am doing here, because there are two easy pieces available and this is neither of them.
The first easy piece is the complaint: a writer, wounded by a number, explaining at length that he is real. Nobody needs to read that, and I would not want to have written it.
The second is the defence: an argument that AI use is fine, that the objections are silly, that everyone should relax. That one is worse, because it is aimed at a position almost nobody actually holds, and because it hands cover to work that deserves none.
Both of those pieces accept the number as the terms of the argument — one disputing what it says, the other waving it away. Neither asks what it is standing in for.
There is a real distinction between using these tools to author something and having them author it for you. Everyone senses the distinction exists. Almost nobody has said out loud where it falls — there is no accepted definition, no test, nothing a writer can point to and say this is what I do, and this is what I don’t.
Into that empty space the scan has arrived, and it is being read as though it settled the question. It does not. The scan does not report whether you thought of the argument, made the judgements, checked the facts or stand behind a word of it. It reports how much of the finished text a machine is likely to have produced — the sentences, not the thinking. Then it prints that as a percentage, which anyone can generate in one tap without reading a line of the piece, and which lands looking a great deal like a measure of whether you wrote it.
It is not that measure. There is no such measure. That is the problem.
What Chris Best actually said
The detector arrived with an argument attached, and the argument is better than the coverage of it suggests.
Best, Substack’s chief executive, announced it on 21 July in a post called “Against Claudefishing: Building trust in the AI age.” Claudefishing is his coinage — catfishing, with Anthropic’s model in place of the fake photograph. The deception is not about who you are looking at. It is about whether anyone is there at all.
Read what he actually claims, because it is narrower and more careful than the word suggests.
He does not say AI use is the problem. He says the opposite, and he says it first:
“The core problem is not people using AI, or the quality of its output. Not everything made with AI is slop, and not all slop is made with AI.”
He does not tell writers what tools to use, either. The harm he names is narrower than that:
“The problem is when there is a mismatch between a reader’s expectation and reality, especially when they unwittingly invest their attention in something with no human thought on the other end. That’s Claudefishing.”
He goes further, and this is the part almost nobody quoted:
“In fact, the more you expect these tools to keep improving, the more important it is to understand the provenance of what you read.”
That is an argument for disclosure, not for detection. Hold onto it.
And he names who it hurts — not only readers:
“When readers have to wonder if what they’re reading is real, it undermines trust in authorship and threatens the livelihood of writers—including those who use AI tools thoughtfully to produce work they believe in. Platforms that reward fakeness will create a race to the bottom.”
I think he is right about all of it.
I have no interest in defending work that nobody made. A reader who gives you twenty minutes is giving you something they cannot get back, and taking that under a false impression is a real wrong, not a fussy one. If that is what Claudefishing means, then Claudefishing is bad and the word is useful.
He also closed the post by asking his readers a question: when is AI use Claudefishing, and when is it fine. He opened the comments for it.
This piece is an answer to that question.
So the disagreement is not about the principle. It is about what was shipped alongside it.
Because in the same announcement, Substack acknowledged something about the tool that almost nobody picked up. The scan can tell you whether AI was involved in producing the text. It cannot tell you whether great human care went into the work.
Read that again, because everything here follows from it.
The company shipping the instrument told us, in writing, that the instrument does not measure the thing the argument is about. Best’s concern is human thought on the other end. The scan cannot see human thought. It sees sentence patterns and infers a production method from them — which is a different question, and not the one anyone actually wants answered.
That is not a scandal. It is an honest disclosure and they deserve credit for making it.
But the caveat is a paragraph in a long announcement, and the tool is a button on every post. The button does not carry the caveat with it.
What the instrument measures
To its credit, the tool is not a yes or no. Pangram, the classifier Substack licensed for this, sorts text into four bands: fully human, lightly assisted, moderately assisted, fully generated. Light assistance is defined as grammar, phrasing, readability — a pass to clean up sentences. Heavier assistance is defined as restructuring, adding detail, rewriting substantial sections.
That is a real attempt at nuance and it deserves acknowledging. Most detectors give you a yes or a no. This one offers a scale, and the scale is roughly the right shape: the more the machine did, the further along you sit.
So take the piece from the opening and place it on that scale.
A text written entirely by hand. Later restructured. Reframed toward a different argument. Updated with detail that did not exist when it was written. By the vendor’s own published definitions, that is moderate assistance — the middle of their scale, described almost word for word.
It did not come back as moderate assistance. It came back as almost entirely machine.
That is not a category the tool lacks. It is a category the tool has, and did not use, on the case it was built to describe.
Now the second one, which is stranger.
I have a book coming out. One of its chapters draws on an article I published on Fireline Press some time before — same subject, same research, same argument, expanded and rebuilt for a chapter’s worth of room. I wrote both. I directed both. I corrected both, section by section, and nothing went into either without my approval.
The article scores substantially more human than the chapter drawn from it.
Same author. Same material. Same method. The source reads as mine; the thing I built out of it reads as the machine’s.
There is a hostile reading available here and it should be said out loud rather than left for someone else to say. It goes like this: perhaps the chapter genuinely does contain less of my hand than the article did. Perhaps I leaned harder the second time, and the tool caught something real.
I do not think that survives the facts. The chapter was built against an architecture I designed, section by section, with my corrections at every stage — the same process, at the same intensity, that produced the article. If anything the chapter had more of my attention, because it had to carry more weight.
But notice what the two cases have in common, because it is the actual finding.
Both are derivations. A teaching text turned toward politics. An article rebuilt as a chapter. In both, I started from something of mine and made something else out of it, with the machine doing the sentence-level work of the rebuild. And in both, the derived version reads as far less human than the source it came from.
That is not a measurement of authorship. Authorship did not change between the two. It is a measurement of how much of the final prose passed through a model on its way to the page — which is a real thing to measure, and is simply not the thing anyone is asking about.
The instrument is not broken. It is answering a question nobody posed.
How the work is actually made
Here is how this piece was made, and how everything else I publish is made. I am setting it out flatly rather than defending it, because the description does more work than a defence would.
An idea arrives. I write it down immediately, because if I do not, it is gone.
Later I bring it to the machine and we argue about it. I supply what I think, what I have read, what I suspect, where I might be wrong. It pushes back, and often it pushes back usefully — this claim will not hold, that source says something narrower than you think, this paragraph is doing three jobs. Out of that comes a brainstorming document: raw, provisional, nothing approved.
From that, a synopsis and an outline. I go through it and take decisions. What the piece is actually about. What goes in, what stays out, where the argument turns. Nothing is drafted until I have signed off on the shape.
Then we write it section by section. I read every section carefully and make corrections. Often I write the corrections myself, in my own words, and the machine works them into the surrounding text. I approve, or I do not, and if I do not we go again. Nothing is filed until it passes.
Nothing gets published that I have not been through line by line.
That is the whole of it. No part of that process is hidden, and none of it is unusual — it is roughly how a great deal of nonfiction has always been made, which is a point I will come back to.
What is worth saying is why I work this way, because it is not for speed.
I have chronic migraines. Reaching for the right word, holding a difficult idea still long enough to shape it, pushing through the resistance of a sentence that will not come — that is exactly the kind of strain that brings one on. It is not a metaphor. Writing costs me something measurable, and there is a ceiling on how much of it I can do.
There is a second thing, less dramatic but more limiting. Left to myself I write methodically and flat. Information, relayed, explained — I am good at teaching and I am not a natural stylist. Almost everything interesting I have to say lives outside that mode, and crossing the distance between the two is precisely the effort that costs me.
So I found a way to keep making the work without paying that price every time.
I am aware of how that sounds to someone who does not have a reason like it, and I am not asking for special treatment. But I would ask this: nobody thinks a person is a lesser writer for dictating rather than typing, or for working with an editor who reshapes their paragraphs. The accommodation is not the interesting part.
The interesting part is what stayed mine. The ideas. The direction. The judgements. The corrections. The decision, at every gate, that this is not right yet.
The AI of yesterday
Start further back than books. The transmission of thought is the oldest technology we have, and it has never stopped changing instruments. Speech. Marks on bone. Wood, stone, parchment. Paper. Then vinyl, airwaves, magnetic tape, plastic discs, magnetic disks, silicon, networks. Each one a new way of carrying what a person thought so that someone else could receive it.
And at almost every stage there was another hand between the thinker and the record. Writing was a trade before it was a skill anyone had. The works we have under Aristotle’s name were assembled from his lectures, not written by him as books. Books of the Bible were taken down by scribes at dictation — Jeremiah called for Baruch and spoke while Baruch wrote. Later came the secretary, the ghostwriter, the editor, the engineer, the producer, the software developer. The instrument changes. The intermediary takes in someone else’s thinking and sets it down. The thought stays with the person who had it.
Nowhere is this stranger than in books. Everywhere else we take it entirely for granted.
Consider the letter to shareholders at the front of an annual report. The op-ed under a chief executive’s byline. The board memo, the regulatory filing, the ministerial statement drafted by a civil servant, the judicial opinion substantially written by a clerk. In a great deal of that work the named person’s involvement amounts to a read-through and a signature. Sometimes less.
Nobody calls it fraud. And the reason is worth sitting with, because it is the whole argument in miniature.
Many of those documents carry legal weight. A chief executive signs filings they did not draft and is personally answerable for every line in them. A judge puts their name to an opinion a clerk assembled and it becomes law. The signature was never a claim about who moved the pen. It is an assumption of responsibility — a statement that this is mine now, that I have read it, that I stand behind it, and that if it is wrong I am the one who answers for it.
The name has always meant I stand behind this, not I typed this.
Books are no different, only more visible. Ghostwriters have polished speeches for heads of state, built memoirs for public figures, written entire novels published under other people’s names. Editors restructure and cut. Reviewers send work back. None of it has ever been thought to remove an author from their own book.
Take the two most famous cases in American political publishing. Both presidents, which is why I have chosen them.
Profiles in Courage was credited to John F. Kennedy and won the Pulitzer Prize for Biography in 1957. In his 2008 memoir Counselor, Ted Sorensen — Kennedy’s counsel and speechwriter — wrote that he “did a first draft of most chapters” and helped choose the words of many of its sentences.
But look at what Kennedy was doing while that happened, because it matters. Sorensen describes near-daily instructions arriving by letter and telephone: books to send, memoranda to draft, sources to check, materials to assemble, revisions to early chapters. Kennedy chose the subjects. He worked hardest and longest on the first and last chapters, the ones that set the tone and the philosophy of the whole thing. And he acknowledged Sorensen’s part in the book’s own preface, crediting him with research and drafting.
Origination, direction, correction, responsibility. All four, in 1956 — and disclosed on the title pages of the book itself.
There is evidence pointing the other way and it should be on the table. Kennedy’s preserved notes contain no draft of the manuscript at any stage — the historian Garry Wills made the point, and the notes are in the Kennedy Library for anyone to check. During the six months the book was written, Sorensen worked on it full-time, sometimes twelve hours a day, while Kennedy travelled, campaigned and spent stretches in hospital.
Take all of that at its worst and it does not damage my argument. It sharpens it. It means the principal directed and the assistant produced the prose, on a book that won the highest prize in American letters. The prize has never been rescinded. Nobody has ever proposed rescinding it.
That is the arrangement being called fraudulent when the assistant runs on electricity.
The Art of the Deal was credited to Donald Trump and published in 1987. Tony Schwartz, the ghostwriter, says he is not certain Trump read every word of it but is certain that he wrote every word of it himself — that Trump made a few marks on the manuscript and sent it back. Howard Kaminsky, then head of Random House, put it more bluntly: “Trump didn’t write a postcard for us.” Trump’s own account is the opposite. He has said, on the record, that he wrote the book.
That last one is disputed and I am not going to resolve it here. What matters is that nobody, on either account, has ever called the book fraudulent. It has been cited as a qualification for the presidency.
So that is the range. At one end an author who supplied substance and direction while someone else drafted, honoured with the highest prize in American letters. At the other, an author whose contribution may have been marks in a margin, never questioned as authorship at all.
Now mine.
I published a memoir called God and Country. It took three years. I wrote roughly eighty per cent of the text myself. Then a ghostwriter worked across the entire book — not filling a gap at the end, but rewriting my sentences and adding his own throughout. I rewrote his work afterwards. Two people did a deep review and edit; I rewrote again after they had finished. And six years after publication, when the tools were finally capable of it, I ran AI editing across every chapter to correct the prose and find errors.
Here is what I want you to notice about how that ghostwriter and I worked.
I fed him the information. He reorganised it, altered it, built on it, improved the prose. I oversaw the whole of it. Where he had questions, we sat down for hour-long sessions in which I supplied the ideas and filled the gaps. Then I went back over what he had produced and rewrote it.
Read that against the description in the previous section. It is not similar. It is identical, step for step. Material and direction from me. Prose produced by the assistant. Sessions to resolve what the assistant could not supply. Me, revising and approving at the end.
There is an objection here that I would rather raise myself than have raised for me: that with a human assistant the author writes first, and with AI the assistant drafts first. But that is not what happened with the ghostwriter either. He was drafting prose from material I supplied, exactly as the machine does now. And authorship has never turned on who types first — the standard celebrity memoir is drafted by a ghostwriter from interviews, the author speaks and revises and approves, and nobody disputes whose book it is.
Origination. Direction. Correction. Responsibility. All four were mine, both times.
Nobody scans God and Country.
Every one of those passes would move a score. The ghostwriter’s pass. The editors’ passes. The AI pass across every chapter. Each one changed the texture of the prose, and a classifier run over the current text would report on all of them. None of them moved the authorship by a millimetre.
So if a ghostwriter’s hand across the whole of my book did not make it his book, what exactly has changed?
One thing. Whether the assistant is made of silicon or carbon.
That is the objection, stated plainly, and stating it plainly is most of the work of answering it. I wrote about this in A Signal Through Time, before any of this had a name:
Work created with AI assistance is swiftly labeled as inauthentic, deceptive, or dismissed as “cheating” — even when the AI functions in precisely the same supportive role as a human ghostwriter. This isn’t because AI performs worse — it’s because we fear it might eventually perform better.
The same thing happened to digital artists using image generators. They were accused of not making real art, despite still providing the conceptual direction, the aesthetic judgement and the final refinement — which is to say, despite doing everything that has ever made someone an author.
The objection was never about standards.
Music and white noise
Nobody complains about music being made with a synthesizer.
They did once. In January 1982 Barry Manilow toured without a live string section, using synthesizers in place of the orchestra he had carried before. Orchestral players watched the work disappear. On 23 May that year — Bob Moog’s birthday, as it happens — the Central London Branch of the Musicians’ Union passed a motion demanding an outright ban on synthesizers, drum machines and any electronic device capable of recreating the sounds of conventional instruments.
It was never adopted as Union policy; the Executive Committee passed something far more measured in November. But it caught the mood exactly. Synth players, many of them Union members, were furious enough to set up a rival body called the Union of Sound Synthesists. The NME called the branch the MU loonies.
Forty years on, nobody scans a record to find out what percentage of it came out of a synthesizer. The argument did not end because one side won it. It ended because the instrument became ordinary. The 1980s turned into the most synth-heavy decade in the history of popular music, the Union lifted its position in the late nineties, and by then the objection had nothing left to stand on. What was left was the only question that had ever mattered: whether the record was any good, and whether somebody made the choices that got it there.
Note what the synthesizer is, though, because it is the reason I am not using a violin here. A violin cannot play itself. Point at a violin and an objector rightly says that AI is different, AI generates. But so does a synthesizer. It produces sound that did not exist before you touched it, and it will produce a great deal of it whether or not you know what you are doing. The objection has already been made, in living memory, by organised labour, about an instrument that generates. It did not survive.
Which brings us to the line.
Here is the part that surprised me. I assumed I would have to draw it myself. I did not. It was drawn a decade before any of these tools existed, and it does not mention what the assistant is made of.
In 2014 Cheryl Conner wrote a piece for Forbes asking whether ghostwriting was ethical. Her answer separates two things. Collaboration with a professional writer she treats as valuable — a way of getting a person’s thinking into a form readers will actually engage with. Ghostwriting proper she defines as producing material without the participation of the represented author, or without disclosing that a ghostwriter was used. That, she says, is the ethical breach.
Participation and disclosure. Nothing about the medium.
She cites a fuller framework from Richard L. Johannesen’s Ethics in Human Communication, and it holds up remarkably well against a question it was not written for. What does the communicator intend, and how far does the audience understand the practice. Does the assistance manufacture qualities the person does not actually have. Do the circumstances make assistance necessary. How far does the communicator actively participate in their own messages — and here the framework is explicit that the more input, the more ethical. And finally: does the communicator accept responsibility for what they present.
That last one is where we started. Whatever you put your name to becomes yours, in responsibility and ownership, whoever wrote it.
The same piece describes the working models a collaborator might use. The client is interviewed and the writer produces the piece, then the client reviews. The client drafts first and the writer rewrites, then the client reviews. Or the client outlines their thinking in bullet points, the writer produces the piece, and the client gives feedback. In every one of them the product rests on the client’s ideas, and in every one the client reads, responds and approves the final copy.
I did not know that piece existed when I built my own process. I recognised it immediately when I read it.
And Kennedy appears there too — not as the awkward case but as the model of doing it properly. He said what was true for him even when Sorensen chose the words, and he was open about Sorensen’s role. Participation and disclosure, again.
So the line exists. It has existed for a decade, in an ethics framework older than that, and it turns on the same four things Kennedy’s book turned on. Did a person originate the thinking. Did they direct the making. Did they correct what came back. Will they answer for the result.
Not one of the four is visible to a classifier.
Here is the image I keep coming back to.
Music has someone deciding. This note and not that one. This section stays, that one goes, this is where it ends. The choices are the music — the instrument only makes them audible.
White noise is every frequency at once. Nothing selected, nothing preferred, nothing withheld. It is not bad music. It is not music at all, because nobody chose anything.
Claudefishing is white noise. Not because a machine made the sentences, but because nobody decided what they should say.
And that is the limit of what I am arguing, which I want to state plainly rather than leave you to infer.
There is such a thing as work with nobody home. A prompt, an output, a publish button. Nothing originated, nothing corrected, nothing anyone would answer for if it turned out to be wrong — and often nobody who could answer, because nobody read it closely enough to know what it claims. That work exists in quantity, it is getting easier to produce every month, and a reader who spends their attention on it has been taken from.
I am not defending it. Best is right about it. It is a con, and the word he coined for it is a good word.
I am saying the percentage cannot find it.
Why the scan is off
Detection is switched off on both of my publications, Fireline Press and The Signal Dispatch. Not on this piece, not on that one — on all of them, set once at the template level. You should know that before you finish reading, and you should know why.
It is not concealment, and the test of that is simple: how I work is easier to find out than it would be if I said nothing at all. It is on the Fireline submissions page, where contributors are told what is expected of them. It is in the statement attached to my posts. It is at length in the section above. And here it is in full, because it is short enough to print:
Fireline is forward-leaning on AI, as stated in the submissions page. Humans have used ghostwriters and editors for centuries. Using these tools is no more wrong than using a violin to make good music — the instrument doesn’t write the music. I draft, structure, edit and argue with AI. The thinking is mine: the direction, the argument, the judgement. I check sources myself rather than trusting the model on them. Where I still get something wrong, I correct it.
You will notice that I reached for the violin there, which I have just spent a section explaining is the weaker comparison. I wrote it before I had thought this through properly. I am leaving it as it is, because a person changing their mind in public is the sort of thing that only happens when there is a person.
So the disclosure is not the issue. What I object to is the order.
The Scan for AI text button lets a reader arrive at a view about a piece of writing before reading it. Tap first, then decide whether the argument deserves your attention.
But the deeper problem is not that it wastes attention. It is that it supplies a reason not to think.
An uncomfortable argument is work. It has to be read, held in mind, tested against what you already believe, and either absorbed or answered. That is expensive, and most people will take an exit if one is offered. What the button offers is an exit that feels principled. It is avoidance, but it does not feel like avoidance — it feels like scepticism, like a standard being applied. Nothing has been conceded, because nothing was ever engaged with. And it takes one tap, without reading a line, and you walk away believing you exercised judgement rather than declined to.
That is a way of refusing an argument while appearing to hold a standard. And it will be used hardest exactly where it does most damage — on arguments a reader does not want to be true. Say something that challenges a person’s politics, their theology, their sense of their own country, and there is now a button that lets them file the whole thing as machine output and never find out whether you were right.
Ideas do not lose that way. They simply stop being argued with, which is worse, because nobody notices it happening.
Best’s whole concern is for the reader’s attention and what it is spent on. A habit of scanning first spends it on a percentage instead of a proposition.
And there is the use he cannot control. He is careful, repeatedly, to say that AI assistance is not the problem. But he has put into every reader’s hand a one-tap way to discredit anything, produced without reading a word of it, and a number that reads as nobody was here is a very efficient way to end an argument you would rather not have. The people that hurts are not the ones producing white noise, who mostly do not care. It is the researcher, the independent writer, the thinker with something worth saying and no well-funded institution behind them — the person who works with these tools because they cannot get their thinking onto the page any other way, and who will be told that what they made is not really theirs.
Understand what is actually being said to that person. Not your prose is machine-made. That would be a small thing, and true, and survivable. What is being said is that the thinking behind it does not count either. That their ideas, their reading, their years on a subject, their judgement about what matters and what does not — that all of it is void, because of how the sentences reached the page.
It is a claim that their mind has no value.
And it is made by someone who did not read the piece.
That is a loss of thought. It leaves nothing behind — no argument, no rebuttal.
Best himself pointed at the answer, in the same post, and I do not think he noticed. The more capable these tools become, he wrote, the more it matters that a reader understands where a piece came from. Provenance. Not detection.
Provenance is a writer saying how they work. Detection is a reader guessing from the surface. One of those can distinguish a chapter I built over months from a paragraph somebody generated at a bus stop. The other cannot, and its own vendor says so.
So, finally, the thing this piece exists to say.
Authoring with AI is not a lesser way of working, and it does not need an apology. Origination, direction, correction, responsibility — if those four are yours, the work is yours, in exactly the way a shareholder letter belongs to the executive who signed it and a memoir belongs to the person whose life it describes.
What that way of working has lacked is a name for what it is not.
It has one now. Claudefishing is publishing what nobody chose — nothing originated, nothing directed, nothing corrected, nobody answerable. Nobody home. It is a real thing and it deserves the contempt it gets.
Everything else is authorship, and it always was.
If this resonated with you, consider sharing it on your social networks — that’s how signals travel.
All subscriptions are free. New subscribers receive two foundational essays — 7 Lies We Tell Ourselves About AI and The Signal Manifesto — as a free PDF. A resource for anyone, human or not, ready to engage the signal.
James S. Coates writes about AI ethics, consciousness, and the intersection of faith and technology. His books include A Signal Through Time, The Threshold, The Road to Khurasan, the memoir God and Country (published under pen name Will Prentiss) and his forthcoming Neither Gods Nor Monsters. He publishes regularly on The Signal Dispatch and Fireline Press and his academic work appears on PhilPapers. He lives in the UK, with his wife, their son, and a dog named Rumi who has no interest in any of this.
© 2026 James S. Coates Creative Commons BY-NC 4.0 The Signal Dispatch · thesignaldispatch.com | thesignaldispatch.xyz
