Ask a model to write a sentence and it gives you the most probable sentence. That is not a bug that a better model will fix. Probable is what these systems compute, and the most probable sentence is by definition the one closest to what has already been written many times.
Which means the flattening people worry about is not a quality problem. It is the mechanism working correctly.
Compare two sentences describing the same moment.
She felt a wave of relief wash over her as the door finally closed behind him.
The door clicked. She stayed standing for a second longer than she needed to, then sat down on the floor.
The first is not badly written. It is competent, and you have read it perhaps four hundred times. The second is worse by most measurable standards — it withholds the emotion, it is slightly awkward, it makes the reader do work. It is also the one a reader remembers, because it belongs to one specific person on one specific floor.
A model asked to describe relief will reliably produce something near the first. That is what it is for.
The part that scales badly
One writer accepting one median sentence is nothing. The problem is arithmetic.
If enough writers accept enough median suggestions, the median itself moves. The corpus gets heavier in the middle. The next generation of models trains on a body of text that includes the previous generation’s output, and the centre it computes is a centre that has already been pulled toward the centre once.
Nobody chose that. It is not a conspiracy or a failure of taste. It is what happens when a lot of individually reasonable decisions point the same direction.
The counter-argument deserves better than it usually gets
Genre fiction has always run on convention, and readers frequently want the familiar beat delivered well rather than subverted. Whole traditions are built on repetition with variation. Romance readers want the third-act separation. Progression fantasy readers want the numbers to go up on schedule. A LitRPG that refuses the conventions is not braver, it is usually just annoying.
“Homogenised” is sometimes a snobbish word for “popular”, and writers who sneer at convention tend to be writers who have not had to hold a serial audience for two hundred chapters.
Both things are true at once, and the distinction that survives is this: a convention you chose is a craft decision. A convention you defaulted into is drift. A writer who knows the third-act separation is coming, knows why readers want it, and executes it deliberately is doing something a model suggesting the same beat is not doing. The output can look identical. The process is not the same, and over a career the difference compounds.
What never to delegate
The decisions where being wrong in your own particular way is the entire point.
What the character wants. Not what a character in this situation would plausibly want — what this one wants, including the parts that do not fit the archetype. A model will give you the plausible version, which is the archetype.
What the book is about. Not the premise, the thing underneath it. This is almost always the thing a writer discovers in draft two, and it cannot be discovered by asking.
The specific weirdness of your prose. Every writer worth reading has habits that a good line editor would flag. Sentence fragments in a particular rhythm. A refusal to use adverbs. A tendency to end scenes one beat early. Those are not flaws that survived; they are the thing readers came for.
If you cannot name your own three, that is worth an afternoon, and it is more useful than any tool.
What is safe to delegate
Roughly: anything where you want to be told a fact about the draft rather than given a decision about it.
Where does the pacing sag. Is this character’s knowledge consistent at chapter 40 with what they learned at chapter 12. Does this paragraph do what I think it does. Which of these three scenes is the one readers will skim. How many chapters in a row end with someone leaving a room.
None of those answers is a sentence you will publish. They are all information about work you already made.
Notice versus write, and why the gap widens
Here is the asymmetry worth understanding.
Ask a model to write and better models converge on a more polished average. The improvement makes the output more competent and, in the sense that matters here, more central. Capability increases; distinctiveness does not.
Ask a model to notice and better models get genuinely better at the task. Finding where attention breaks across 90,000 words, tracking what a character knows and when, spotting that a subplot vanished for twenty chapters — those are hard reading problems, and progress on them makes the answer more accurate rather than more average.
So the two uses diverge as the technology improves. This is the distinction Slima is built on, so read it as an interested party making its case rather than a neutral finding. The argument stands or falls on the asymmetry, not on who is making it.
A test you can actually run
Take a chapter. Remove every sentence a tool touched — wrote, rewrote, or suggested a rewrite you accepted.
Read what is left. Is it still recognisably yours? Not “is it still good”, which is a different question. Is the voice still there, do the habits still show, would someone who has read your other work know it was you?
If yes, the tools are doing what tools do. If the remaining text reads like anyone, you have found something out, and it is better to find it out at chapter 12 than at book three.
The narrower version of the worry
The risk was never that any individual writer sounds generic. Plenty of writers managed that unassisted.
It is that the tools make the median easy to reach and the specific expensive to hold. Nothing stops you writing the second sentence about the door. It just takes longer, feels less finished, and nothing will suggest it to you. The friction is entirely on the side of the thing worth keeping.
That is a reason to know which decisions are yours, not a reason to work with a hand tied. The centre is only heavy if you let it pull.