A thermostat can tell you the room is cold. It’s never been cold.
Ask a model to write something moving and it produces a passage where the emotions are correctly identified, appropriately escalated and accurately named. It’s competent. It’s also inert, and most readers can feel that within a paragraph even when they cannot explain it.
The gap isn’t vocabulary. It’s that feeling on a page comes from a writer having something at stake, and stake is a fact about the writer.
What does emotion on a page consist of?
A cost the writer paid, visible in what they chose to include.
The signal is rarely the emotional language. It’s a detail somebody would only include if it mattered to them, or a sentence that stops short because going further was difficult, or a refusal to soften something that a careful writer would have softened.
When I wrote about a marriage ending in The Relationship BULLSHIT Men Fall For, the sentence that does the work is that I won’t pretend I was devastated. That’s four words admitting something unflattering, and it works because there was a reason not to write it.
A model has no reason not to write anything. Every sentence available to it costs the same, which is nothing, and that uniformity is what a reader is detecting.
The book on this: The Relationship BULLSHIT Men Fall For contains the sentence this piece is about, and I nearly cut it.
Why does the generated version feel padded?
Because it reaches for intensity where a person would reach for a detail.
Asked for grief, a model supplies the recognized markers of grief. Asked the same thing, a person who has been through it usually produces something oblique and specific: the abundance of food at every meal, a shirt kept in a drawer, a case of pins put back on a shelf.
The oblique version carries more because the reader has to do the arithmetic, and the arithmetic is the feeling. The direct version tells them what to feel and they decline.
This is also why AI-drafted emotional writing tends to run long. Without a specific detail to rest on, it substitutes accumulation, and accumulation of correctly named emotion is exactly the texture readers describe as fake.
The book on this: My Life in Crazytown is 19 years of material where the restraint is the technique and the naming of feelings would have killed it.
Can you fix it in editing?
Partly, and the fix is always subtraction plus one true thing.
Cut every sentence that names an emotion. Most drafts survive this without losing anything, which tells you what those sentences were doing. Then add one concrete detail from what happened, and the passage usually works.
What you cannot do is edit stake into a passage where there’s none. If the writer had nothing invested, no amount of revision produces the thing, and the conclusion is that this section should be shorter and plainer instead of emotional at all.
A flat factual paragraph is respectable. A paragraph performing feeling it doesn’t have is the worst outcome available.
Can a business book handle emotion well?
It’s where most of them fail, though nobody frames it that way.
Business books rarely attempt emotion directly. What they lack is the smaller version: any indication that the author minded. A chapter about a failed project written without the slightest sense that it was unpleasant reads as a case study instead of experience.
The fix isn’t a moving passage. It’s one sentence admitting that the quarter was miserable, or that you were wrong and it cost somebody their job. That single admission changes the temperature of a chapter and it takes about eleven words.
A model won’t write that sentence, because it doesn’t know it happened and wouldn’t consider it necessary.
The book on this: The Ethical Workplace is a business book that works because one sentence in it admits the author minded.
How do you tell if a passage has real emotion?
Look for what the writer risked including.
Every page with real feeling on it’s at least one line the author could reasonably have cut for comfort. An admission, an unflattering detail, a conclusion that doesn’t resolve nicely.
Read a chapter and ask which sentence was difficult to leave in. If there’s not one, the chapter is informational, which may be correct for that chapter and is worth knowing on purpose instead of by accident.
The test works on your own manuscript and it works on anything you’re assessing, whether or not a machine was involved.
Where does that leave AI as a writing tool?
Useful everywhere the temperature is meant to be neutral.
Explanation, procedure, background, structure. Large parts of any nonfiction book are supposed to be flat, and a model is very good at flat and never gets bored, which a human writer does.
What it shouldn’t be handed is the two or three passages in a book where the reader finds out what it cost you. Those are load-bearing, they’re short, and they’re the reason anybody finishes.
The rest of this series covers the other gaps, including why AI never writes in your voice and when to bring a human back into the work. The AI and Writing Hub collects them, and my ghostwriting service spends much of its time finding the sentence a client nearly left out.