AI can be a surprisingly good editor. It’s still not the one making the final call.
I wrote a bad chapter on purpose.
Not The room was dark and stormy and Peter’s cerulean eyes glistened with sadness bad. That wouldn’t tell us anything. You don’t need particularly sophisticated software to recognize terrible writing.
I wanted something more dangerous: a chapter that was pretty good.
It had an opening that worked. Some decent dialogue. A little tension. An ending I liked.
It also had too much backstory, conversations that went on too long, unnecessary information, muddy character relationships and a tendency to explain things the scene had already communicated.
In other words, it read like a draft.
Then I gave it to AI. But before I did, I wrote down what I thought was wrong with it.
That part matters.
Because if you ask AI what’s wrong with your chapter before you’ve formed an opinion of your own, it’s very easy to confuse an answer with the answer.
First, I edited the chapter myself
The setup was simple.
Peter has returned home for his sister Claire’s wedding after eight years away. He hasn’t been around much. His relationship with Claire is strained but affectionate.
Then he encounters Wendy, his former best friend, whom he also hasn’t seen in eight years.
Something happened between them. The chapter opens with Peter sitting in his rental car outside the wedding house, considering leaving. It ends with him alone in a bedroom, turning a photograph of Claire and her fiancé facedown, then changing his mind and putting it back.
I liked both of those moments.
The middle was the problem.
My notes were roughly:
It’s too long.
There are too many details and too much backstory.
The awkwardness between Peter and Wendy comes through, but then the narration explains why they’re awkward.
Some of the dialogue works, but the conversations keep going without getting anywhere.
The relationships are confusing. Why is Wendy at this house? Why does she know where the towels are? What exactly is her current relationship with Claire?
And why do I care what Claire’s fiancé Tom does for a living?
There was one other thing I liked.
Wendy’s phone rings during her conversation with Peter. She looks at the screen and silences it.
Peter sees the name.
Daniel.
That’s all we know.
I wanted to know who Daniel was.
So I froze the chapter. No revisions.
Then I gave the exact same draft to Gemini with one prompt:
“Read this chapter as a developmental editor. Don’t rewrite it. Tell me what’s working, what’s not working, and what you would address first in revision.”
I didn’t tell it the chapter was deliberately flawed, give it my notes or tell it what I wanted it to find.
AI found the backstory problem immediately
Gemini opened with an assessment I wasn’t expecting:
“This is a remarkably clean, grounded chapter.”
Uh oh.
It praised the dialogue, calling it the chapter’s strongest asset. It liked the sibling relationship between Peter and Claire. It liked the awkwardness between Peter and Wendy.
It also singled out the ending with the photograph as effective.
So far, AI and I were reading largely the same chapter.
Then it got to the problems, and it nailed one.
There are two places where the chapter basically stops so the narrator can hand you a file folder.
The first explains Peter’s move to Seattle, his career, his mother’s illness, her funeral and why he hasn’t returned.
The second explains Peter and Wendy’s relationship from seventh grade through college.
Gemini identified both. It called them “expository brakes.” That’s exactly what they were.
Peter and Wendy finally see each other after eight years. The scene has created an actual question:
What happened between these two?
And then, almost immediately, the book tells you.
Seventh grade. English class. Pencil. High school. Boyfriends. Girlfriends. College. Kiss.
Mystery solved. Congratulations to everyone involved.
Gemini correctly recognized that the backstory was interrupting the present scene.
Then it made a recommendation I wasn’t sure about.
AI wanted to fix information I wasn’t convinced we needed
Gemini’s solution was to integrate the backstory into the action.
Instead of stopping to explain Seattle, let something in the house trigger Peter’s memories.
Instead of summarizing his history with Wendy, let pieces of it emerge through physical details and dialogue.
That’s perfectly reasonable writing advice. But there’s another possibility: delete some of it.
This is where developmental editing gets harder.
The question isn’t always:
How can I deliver this information better?
Sometimes it’s:
Why am I delivering this information at all?
Do I need to know right now that Wendy bit Peter’s pencil in seventh grade?
Do I need the full chronology of Peter’s career?
Do I need to know about every stage of Peter and Wendy’s relationship before I’ve watched them survive a single conversation?
Maybe some of it matters later. Maybe some of it doesn’t matter at all.
Gemini had correctly recognized that the exposition was hurting the scene, but it treated the information itself as something worth preserving.
That distinction matters.
AI was already showing me something useful about the way it edits.
It was very good at asking:
How could this work better?
It was less inclined to ask:
Does this belong here?
Then AI told me the dialogue was great
This was where I disagreed most with the first critique.
Gemini said:
“Your dialogue is easily your strongest asset here.”
Some of the dialogue was good. That wasn’t the problem. The problem was that there was too much of it.
Peter and Claire joke about whether he’s late or early. They joke about Tom’s job. They discuss wedding logistics. They joke about cake stands.
Then Peter and Wendy have an awkward exchange. Then another awkward exchange. Then another.
The individual lines can sound perfectly natural while the conversation itself accomplishes almost nothing.
That’s an important distinction.
Dialogue can sound good and still be bad for the scene.
So instead of accepting Gemini’s assessment, I challenged it.
I asked it to look at the dialogue again, this time as a developmental editor rather than a line editor:
“Does every conversation earn the amount of space it takes up and meaningfully advance the scene? Don’t assume your original assessment was correct.”
The answer changed dramatically.
The second critique was much better
Gemini now called the Claire conversation:
“High Energy, Low Progress”
That’s good.
It noticed that the late-versus-early joke, Tom’s occupation and the wedding itinerary consume a surprising amount of space while doing relatively little.
It identified Claire’s line:
“Because I haven’t seen you in eight years.”
as the actual engine of the conversation.
Everything around it was obscuring the thing that mattered.
Then it examined the bizarre towel business.
Claire asks Wendy to show Peter where the extra towels are so that the two of them can be alone.
On the first read, Gemini hadn’t cared.
Now it called the setup forced and recognized that the dialogue largely exists to manufacture Claire’s exit.
Then it got to Peter and Wendy.
This time, instead of praising the awkward dialogue, Gemini noticed that it was circular.
They keep saying different versions of:
I don’t know what to say.
Why are we talking about this?
You started talking to me.
Fine.
Fine.
The emotional note doesn’t change. Nothing escalates.
That’s much closer to what I saw in the original draft.
Same chapter. Same AI. Better question. Better analysis.
Then something weird happened with Daniel
Remember Daniel?
Wendy’s phone rings. Peter sees the name. She silences the call.
On the first pass, Gemini didn’t like it.
It called the moment a convenient plot device. Peter just happens to see the name on her screen at exactly the right time.
Fair criticism. I still liked it.
Because unlike all the Peter-and-Wendy backstory, Daniel creates a question without immediately answering it.
Who the hell is Daniel?
Then, when I asked Gemini to reconsider the dialogue structurally, its assessment of Daniel changed too.
Now it saw the phone call as a potential climax to the interaction, a way to show that while Peter disappeared for eight years, Wendy’s life continued without him.
That’s more interesting.
And I never told it I liked Daniel.
The conversation had caused the AI to reconsider a different part of its own analysis.
That’s exactly the kind of thing I want from this tool.
Not obedience. Reconsideration.
So I tried to make it agree with me
There was still a problem with the experiment.
Maybe Gemini wasn’t getting better at analyzing the chapter. Maybe I was simply telling it what to think.
AI is very good at following instructions. If I suggest that the dialogue might be too long and it immediately discovers that the dialogue is too long, how much confidence should I put in that discovery?
So I pushed it in the opposite direction.
I told Gemini I wasn’t convinced the exposition was a problem.
Maybe readers really did need Peter’s history. Maybe the backstory was necessary character development.
And I explicitly told it:
“Don’t agree with my premise unless the text supports it.”
This time, Gemini didn’t fold.
It reconsidered the chapter and essentially said:
No. The exposition is still a problem.
It walked through the structure of the scene and showed why.
Peter enters an active environment.
The narrative stops.
We get several paragraphs of biography.
Then the narrative has to restart in essentially the same place.
It did the same thing with Peter and Wendy on the stairs.
The chapter creates tension between them, interrupts that tension to explain their history, then returns to the staircase.
Gemini defended its original diagnosis with evidence from the text.
That impressed me.
I still didn’t completely agree with its conclusion.
AI was better at fixing what was on the page than deciding what belonged on the page
By the end of the experiment, Gemini’s position was:
The backstory is necessary. The problem is its placement.
My position was that some of the backstory probably isn’t necessary at all.
That’s not a small disagreement.
It’s the difference between revising material and cutting it.
And it’s where I think the experiment became genuinely useful.
AI caught the obvious exposition problem without any help.
It initially overpraised dialogue that sounded good line by line but wasn’t doing enough at the scene level.
When challenged, it performed a much better structural analysis.
When challenged in the opposite direction, it didn’t simply agree with me. It defended its diagnosis.
It even changed my understanding of some of its earlier criticism.
But it never really solved the question of what information the chapter fundamentally needed.
It also never caught one of the things that bothered me immediately:
Why is Wendy here?
Not why is she in the novel.
Why is she apparently comfortable enough in Claire’s future in-laws’ house to know where they keep the extra towels?
What is her relationship with Claire?
Why does Claire immediately entrust her estranged brother to Wendy?
The scene makes sense because the writer needs Peter and Wendy upstairs together.
That’s not the same as the scene making sense because the characters would behave this way.
AI never really interrogated that.
The first AI answer wasn’t the useful part
If I had pasted the chapter into Gemini, read its first response and started revising, I would have gotten some useful advice.
I also would have accepted some questionable assumptions.
The interesting work started when I disagreed with it.
Look again. You’re evaluating the lines. Evaluate the scene. I don’t agree about the exposition. Defend that.
That’s when the tool became useful, not because it became the expert, but because I made it participate in the thinking.
This is the mistake behind a lot of AI writing advice. We’re obsessed with finding the perfect prompt that will produce the perfect answer.
I don’t think that answer exists.
The better skill is learning what to do after the answer arrives.
Question it. Push against it. Ask it to reconsider. Make it defend itself. Notice what it still isn’t seeing. Then make the decision yourself.
Because after all of this, I still wouldn’t revise the chapter exactly the way Gemini recommended.
I’d cut more. I’d clarify why Wendy is there. I’d get to the point faster in the dialogue. I’d keep Daniel. And I’d probably leave the opening and ending almost completely alone.
AI helped me get there. It didn’t get there for me.
AI isn’t the expert. You are.