TL;DR: The one risky, specific thing only a person who lived the work would know to say is exactly what AI leaves out. It’s trained on averages, so it hits the common arguments and lands the common conclusions, and the result reads like a brochure. Rewriting a few sentences doesn’t touch a problem that structural. The fix is to start from your real material in your real voice and write it yourself, and that starting point is how voice capture works.
Here’s an AI paragraph on running a customer service operation:
Effective customer service operations require a strategic balance of technology, well-trained staff, and clear processes. Leading organizations use data-driven insights to optimize the customer journey, ensuring consistent quality across all touchpoints. By helping frontline agents with the right tools and authority, companies can drive both customer satisfaction and operational efficiency, creating sustainable competitive advantages in today’s dynamic marketplace.
Here’s a paragraph by a human who has run a customer service operation:
The first thing you learn running a customer service team is that the worst hour of the week is Monday morning at 9, because that’s when every weekend’s worth of broken purchases lands at once, and your best agents are still on coffee. The second thing you learn is that the agents who solve the hardest problems are the ones who don’t read the script, and your job is to figure out who they are before HR fires them for not reading the script.
Both paragraphs are technically correct. One is dead.
The deadness has a name. Shallowness in AI prose breaks more attempted books, articles, and reports than any other failure mode, and it can’t be edited out. It does the most damage when the subject is somebody’s own life or business. That person had the one detail that would’ve made the page worth reading, and they handed the job to a machine that could never know it.
What is AI shallowness?
Shallowness in AI prose is not a grammar problem or a tone problem. You can read AI output and find no single sentence you’d cut. Every sentence is technically fine. Every paragraph holds together. The piece reads like a piece somebody would write.
You can read AI output and find no specific sentence you’d cut.Share on X
The problem is what isn’t there.
The AI paragraph on customer service tells you nothing you didn’t already know.
The strategic balance of technology and people. The data-driven insights. The customer journey. The frontline agents with the right tools. None of it is wrong. None of it is specific. None of it is the thing only somebody who ran the operation would know to say.
Shallowness is the absence of the specific, risky, slightly controversial detail a working writer puts in because they earned it. Monday morning at 9. The agents who don’t read the script. The HR problem of keeping the good ones. None of those show up in an AI paragraph, because none of them is average. You only know them if you sat in the chair.
Why is AI prose shallow?
The mechanism is structural to how large language models work. It’s not a flaw better prompting or more training can fix.
A language model produces text by predicting the most likely next word, given everything that came before. Across millions of training examples on a subject, the “most likely” word is the average by definition. The word most writers would use in that position. The phrase most writing on the topic includes. The conclusion most pieces in the genre reach.
The system has no concept of choosing. It’s a probability machine producing the highest-probability output, and across millions of training examples, the highest-probability output is the consensus. Consensus is the average of what everybody already said, and a book built out of that is shallow before anyone reads a word of it.
This is why prompting doesn’t fix it. Ask the AI to “write in a punchy specific voice with concrete details” and it produces text that looks like text with concrete details. The details will be the average details people write when they’re trying to look specific. They won’t be the ones only somebody who lived the work would know. The system can fake the look of specificity. It can’t produce specificity, because specificity comes from lived experience, and the system hasn’t lived anything.
Why can’t editing fix AI shallowness?
People try. I see it constantly with clients who drafted something with AI and want to “warm it up” before publication. The instinct is to read the draft and swap the most obviously average sentences for sentences with more voice.
It doesn’t work, because the shallowness is in the structure underneath them. The AI organized the content the way AI organizes content. It hit the common points in the common order. It made the common arguments. It used the standard examples. It reached the standard conclusion. Rewriting a few sentences leaves all of that in place.
I don’t blame people for trying the cheap fix first. I blame the course sellers and prompt gurus who promise that one more pass will give an AI draft a soul. It won’t, and the author finds out after their name is already on the cover.
I covered this in a piece on AI-written books that get sent to me for review. The author tries to save the chapters by adding a story here, a joke there, a personal anecdote in the middle. The result beats raw machine output and is still no good. It reads, in the words I used there, like a hollow thing wearing a couple of human accessories, because that’s what it is. The full version is in You Used AI and It Shows.
The only thing that works is rebuilding from real material in your real voice, with the AI draft as a scaffold you throw away. That’s a rewrite from scratch. People hate hearing it, and I’d be lying to them if I called it editing.
What do readers notice about shallow AI prose?
From the podcast
Trevor Sumner makes the same point from the supply side, on Leaders and Their Stories:
One of the things about Gen AI is it’s going to output a tremendous amount of content. And the question is, how good is that content?The volume at the bottom explodes. The value at the top goes up, because there’s suddenly so much more to be better than.
The reader can’t always name what they’re noticing. They can feel it. The piece reads competent and they put it down without finishing. The book is about a subject they care about and they never recommend it. The article lands in the tab pile and gets closed without action. That quiet exit costs a writer more than a bad review would. A bad review at least tells you something. A reader who drifts off never mentions the book to anyone, never hires the author, never comes back, and the author never learns why.
The only thing that works is rebuilding from real material in your real voice, with the AI draft as a scaffold you throw away.
Ask them why and they shrug. They might say “it didn’t grab me” or “I wasn’t really into it.” What they sensed was the absence of the specific. Somewhere below words, they expected a person who knew the subject to say something surprising, to commit to a position, to risk being wrong, to include a detail only that person would have included. The absence of all that produces the disengagement. The reader didn’t notice the AI. They noticed the work was hollow and stopped reading.
Most people writing about AI prose get the cause backward. The usual argument, “readers can tell when AI wrote something,” overstates it. What readers can tell is that something is hollow, and when they go looking for why, they find the AI. The hollowness is the signal, and the AI is the explanation they land on second.
The implication is dangerous, because it lets writers believe that hiding the AI better will save them.
It won’t. The reader responds the same way to hollowness whether it came from AI, from rushed human writing, from a writer who didn’t know the subject, or from any other source of vacancy on the page. I’ve got no use for the tools that promise to make AI text undetectable. They sell a disguise to people who need substance, and the reader still walks away.
What working writers do
Working writers, the ones whose books and essays land, are doing something specific that AI can’t do.
They’re including the one detail only they would have included.
The 71-year-old client I work with on a memoir is doing this on every page. He includes the street his uncle worked on in 1972. How his mother kept a glass jar full of buttons and no two were ever the same color. The argument with his father that lasted three years and ended over a single sentence neither of them remembered later but he can still quote.
Every page has at least one detail like that, and the details make the page worth reading.
The AI would have written the same chapter as “a complicated relationship with my father that shaped my early adulthood, marked by both warmth and disagreement.”
Technically true, completely useless. The human writes “my father said one thing in 1976 that I haven’t been able to get past and I’m seventy-one and probably never will.” Specific, risky, says something only this writer could have said. The full profile of how he does this is in The 71-Year-Old Memoirist Who Uses AI Better Than You Do.
That’s the skill: including the things only you would have included. The machine can’t fake it, ever, because it has nothing only it would have included. I’d trade a whole chapter of smooth summary for one detail like that button jar, and so would any reader who picked up the book.
My take
No disguise saves a hollow page. Readers walk away from vacancy wherever it came from, and a tool that makes AI text undetectable can’t add substance. Write from your own material. Put in the street name, the jar of buttons, the one line from an old argument you can still quote word for word. One detail like that does more work than a chapter of smooth summary. The machine can’t fake it, because it has nothing only it would have included.
The mechanism, named
Shallowness is the structural inability of a language model to produce specificity, because specificity comes from lived experience, and the language model has lived nothing.
You can’t prompt around it. You can’t edit around it. You can’t sprinkle around it. You can’t disclose your way around it. The only fix is starting from your real material in your real voice, the work AI was supposed to save you from doing, and the work that produces something worth reading.
This is the failure mode The Death of Thinking is about at scale. What happens to a profession, an industry, a culture, when the loudest output in the room is the structural average of everything written before it, dressed up in clean grammar, with nothing of any specific person in it. The answer isn’t encouraging, and it’s the answer the next decade will produce unless individual writers, individually, keep doing the work the machine can’t do.
That work is yours, and I think handing it to a machine is the most expensive shortcut a writer can take. You save a few weeks and lose the only reason anybody would read the thing. Start with the detail only you would have included, and don’t let any tool talk you out of it.
Shallowness is the price of letting a machine think for you. The Death of Thinking is my book on that cost and how it builds without anyone noticing.
Frequently Asked Questions
- The Surprising Power of Next Word Prediction: Large Language Models Explained, Part 1, Center for Security and Emerging Technology, Georgetown University
