Ask a model to write a chapter on your specialism and it’ll produce something competent in seconds.
Read it carefully and you’ll notice something specific. It’s the average of your field. Every claim is one most practitioners would accept, every emphasis is the standard emphasis, and nothing in it would surprise anybody who has worked in the area for five years.
That’s the tool working as designed. It has no way to know which received ideas you’ve stopped believing, or why, so the average is all it can give you.
This is the problem I care most about when somebody asks whether they should write their book with AI. The tool is fine for plenty of jobs. The trouble starts when an author hands it the part only they can supply, because the book comes out sounding like everybody and the name on the cover is the only thing in it that’s theirs.
What does a model know about your subject?
The consensus, weighted by how often it’s been written down.
It knows what the textbooks say, what the popular articles say, and what the widely repeated framings are. On most subjects that’s a great deal of accurate information and it’s useful for orientation.
What it doesn’t contain is the private half of your expertise. The thing you were taught that turned out to be wrong in practice. The rule everybody repeats that you stopped applying in 2011 because you watched it fail. The exception you’ve seen forty times that nobody has written up because it’s inconvenient.
That private half is your entire competitive position as an author. It’s the only reason somebody should read your book instead of the three that already exist.
I’d go further. An author who publishes the average version has spent months and real money telling readers what they could have found on the first page of a search, and then signed it.
The book on this: The Birth of the Augmented Human is 31 chapters on using these tools as an amplifier, with a companion volume on what happens when the judgment goes across too.
Why does the average version feel so plausible?
Because consensus is usually correct, and correct is what makes it dangerous.
If the draft were wrong you’d catch it. The problem is that it’s right, in the way a good textbook is right, and reading your own subject stated correctly produces a pleasant sensation of the work being done.
Then it ships, and it reads like every other book in the category, and the author can’t work out why nobody responded to it.
More and more of the books people bring me to fix began exactly this way. They wrote it with AI, it came out as crap, and now they want it rescued. Rescuing it means sitting the author down and interviewing the private half out of them, the same work they’d hoped the tool would spare them.
The books in my library that do anything all contain something a competent generalist wouldn’t have written. It came from somebody who had been there, and being contrarian had nothing to do with it.
The book on this: The Death of Thinking is that companion volume, and it argues the case this piece is making at book length.
Can you tell the model about yourself?
Up to a point, and the ceiling is lower than the enthusiasm suggests.
The books in my library that do anything all contain something a competent generalist wouldn’t have written.
You can supply background, examples, and preferences, and that improves the output. What you can’t transfer is the thirty years of pattern recognition that makes you notice one thing in a situation and dismiss another. Those are weightings you can’t list, and a prompt can only carry what you can list.
Try it directly. Write down the ten things you know about your field that aren’t in the books. Most people manage four and then stall, and the stall is the point. The rest surfaces when a situation triggers it, so it comes out in conversation and interview. A good question from another person triggers it far more reliably than a blank prompt box ever will.
The book on this: Family Cybersecurity contains the private half of thirty-three years in enterprise technology, the part no model has.
What does AI not knowing you mean in practice?
Use the tool where the average is fine and stay out of its way where it’s not.
The average version is perfectly adequate for the connective material. Background a reader needs but doesn’t care about. Definitions. The bridge between two chapters. A first pass at a structure you’ll then rearrange.
It’s not adequate for the argument, the examples, or anything that establishes why you’re worth reading. Those parts have to come out of you, and the tool can’t tell you which is which, because it doesn’t know which parts are the average.
A reasonable division: it drafts, you supply, you decide. The moment the decision moves across is the moment the book becomes the aggregate.
| Part of the book | Average version adequate? | Who supplies it |
|---|---|---|
| Background the reader needs | Yes | The tool can draft it |
| Definitions | Yes | The tool can draft it |
| Bridges between chapters | Yes | The tool can draft it |
| First pass at a structure | Yes | The tool drafts, you rearrange |
| The argument | No | You |
| The examples | No | You |
| Why you’re worth reading | No | You |
| Final decisions | No | You decide |
My take
Models hand you the average of your field, and your book can’t afford to be average. More and more of the manuscripts people bring me to fix started this way. They read correctly and they read like everything else on the shelf. Write down the ten things you know that your colleagues would argue with. That list is your book. Use the tool for structure if you want, but the private half has to come out of you.
How do you tell if it has happened to your manuscript?
Give a chapter to somebody senior in your field and ask what surprised them.
If the answer is nothing, the chapter is the average, whether or not a machine was involved. Plenty of manuscripts written entirely by hand have the same problem, because the author wrote what they thought a book on the subject should contain instead of what they know. The second test is subtraction. Remove your name from the chapter. Would anybody in your field be able to guess who wrote it? If not, there’s nothing in there that only you could have supplied. That test is uncomfortable and it takes about a minute.
I’d run it before a manuscript goes anywhere near a publisher. Authors hate hearing that anyone could have written a chapter, but hearing it from a colleague costs a minute, and hearing it from readers after launch costs the book its one chance.
The rest of this series covers the neighboring gaps: AI never writes in your voice, and the labor split that works on a book. The AI and Writing Hub collects the rest, and my ghostwriting service is largely the business of getting the private half of somebody’s expertise onto a page, the half no model has.
Experts with decades in a field who publish something a model assembled from everybody else’s writing waste their best asset, because the readers who needed what that expert knew never get it. The tool is fine for the connective tissue. Hand it the argument and you’ve thrown away the only reason anybody would buy a book with your name on it.
No model can hold your judgment for you. The Death of Thinking is my book on what outsourcing judgment costs, tracing how reasonable shortcuts pile up until the thinking you relied on is gone when you need it.
