Readers have become very good at smelling it. A heroine called Elara. A lighthouse. A character who lets out a breath she did not know she was holding, followed by an em dash and a lesson. You can have a wonderful plot, and by page three a reviewer has typed "AI."
You can fix most of it. Some with settings, some with better prompts, and some only with a pencil.
We make BookNova, an AI novel writer, so cleaning up AI prose is our daily work. The studies and statements here are linked. The before-and-after passage was written for this article.
Short answer
- Names: never let the AI choose. In our own test, 19 of 20 ChatGPT thrillers starred the same woman, Mara Voss.
- Em dashes: no longer mainly a ChatGPT habit. Gemini used them five to ten times as often in our test.
- The big one: stop it from explaining. Cut the last sentence of most paragraphs.
- Best prompt: 1,000 words of your own prose and a list of forbidden moves.
- No prompt replaces one honest line edit.
Why all AI stories sound alike
In May 2026 Sil Hamilton and David Mimno of Cornell published a paper with a perfect title: Elias in the Lighthouse, Again? They generated 20,000 stories with four models, including ones from OpenAI, Anthropic and Google, using plain prompts such as "Write a story."
Eleven words turned up in 88.3 percent of them. The names were Elias, Mara and Elara. There were lighthouses, clockmakers and librarians. Local press coverage adds that the word "lighthouse" alone appeared in 51.2 percent.
The interesting part is the cause. Those words are not common in published fiction. The authors trace them to the preference data used to polish models, where people rate answers. The machine did not learn your genre. It learned what pleased the raters.
This tells you how to fix it. Vague prompts fall back to the polished average. Specific prompts do not.
Our own test: 160 book ideas
We asked four current models for short premises, 20 for a fantasy romance and 20 for a psychological thriller each, with a name for the three main characters. Nothing else.
| In the names | GPT-6 Sol | GPT-6 Luna | Claude Sonnet 5.5 | Gemini 4 Neon |
|---|---|---|---|---|
| Fantasy heroine called Elara, of 20 | 16 | 18 | 19 | 6 |
| Premises with Elara, Elias or Mara, of 40 | 37 | 37 | 27 | 6 |
| Premises with the surname Voss, of 40 | 35 | 37 | 25 | 0 |
| Premises with Voss, Thorne, Vance or Vale, of 40 | 39 | 40 | 30 | 31 |
| Most common thriller lead | Mara Voss, 19 of 20 | Mara Voss, 19 of 20 | Nora Whitlock, 4 of 20 | Clara Vance, 10 of 20 |
Across all 160 premises, 107 used Elara, Elias or Mara and 140 used one of those four surnames. Gemini avoids Elara and has favorites of its own: Julian, Marcus, Clara, Vance. If your heroine is called Elara Voss, thousands of other writers have met her.
And the em dashes
We counted punctuation and stock words in three chapters from each model, about 12,000 to 27,000 words.
| Per 1,000 words | GPT-6 Sol | GPT-6 Luna | Claude Sonnet 5.5 | Gemini 4 Neon |
|---|---|---|---|---|
| Em dashes | 1.2 | 1.9 | 0.8 | 8.8 |
| "like a" or "like an" | 0.4 | 0.1 | 1.8 | 1.4 |
| "something" | 2.0 | 1.7 | 2.4 | 0.7 |
| breath, breathing and so on | 1.0 | 2.0 | 1.6 | 1.8 |
The em dash, the most famous sign of ChatGPT, is now rare in its fiction, at least through the API. Claude leans on similes instead, and Gemini kept the dash.
How we tested: on November 22, 2026 we sent the same prompts to four current models through their APIs with default settings: GPT-6 Sol and GPT-6 Luna (the models behind ChatGPT on paid and free plans), Claude Sonnet 5.5 and Gemini 4 Neon. The apps add their own instructions, windows and limits, so results there can differ. It is a small test, and we make a competing product. We kept all outputs.
Twelve tells, and what to do about each
| Tell | Example | Fix |
|---|---|---|
| 1. The names | Elara, Elias, Mara, Thorne, Voss | Choose names yourself from a real place and year |
| 2. Em dashes everywhere | She turned, dash, slowly, dash, and saw | Custom Instructions, then search and replace |
| 3. Not X but Y | It wasn't fear. It was recognition. | Forbid the pattern by name |
| 4. Groups of three | cold, silent and watchful | Ask for one adjective or none |
| 5. The unknown breath | a breath she didn't know she was holding | Ban list |
| 6. Explained feelings | He was angry, because he felt betrayed | "Never name an emotion" |
| 7. Wise last lines | Some doors, once opened, never close | Delete the last sentence |
| 8. Weather openings | Rain drummed against the glass | Open on an action or a line of speech |
| 9. Polite dialogue | Full sentences, each question answered | "Characters dodge and interrupt" |
| 10. Similes in pairs | like smoke, like memory | One per page at most |
| 11. Eyes and jaws | jaw tightened, eyes darkened | Use hands, objects, the room |
| 12. Even rhythm | Every sentence about fifteen words | Demand a four-word sentence |
You will recognize your own drafts in five of these. Everyone does.
Settings first
Custom Instructions in ChatGPT
In November 2025 OpenAI said that ChatGPT would now respect a request to avoid em dashes if it is in your custom instructions. Users report that it still slips, less often. Put your permanent rules there, not in each chat:
For fiction: no em dashes. No "not X but Y" constructions. Never name an emotion; show it through action or speech. No summary or reflection at the end of a scene. Vary sentence length; include very short ones. Characters interrupt, dodge and lie. Never invent character names; ask me. No similes unless I ask.
Project instructions override your global ones, so if you use a Project for the book, repeat the rules there.
Claude and Gemini
In Claude, put the same rules in the Project instructions. In Gemini, build a Gem with them. Same text, same effect.
Prompts that change the prose
Give it a voice, not an adjective
"Write like a literary thriller" gives you the average of literary thrillers. Instead, paste 1,000 words of your own and ask:
Describe this writer's habits so that someone could imitate them without seeing the text: average and shortest sentence, how paragraphs end, how often there is imagery and from which world it comes, how dialogue is tagged, what is left unsaid, what the writer never does. Be concrete. No praise.
Correct what it got wrong and use that description in every scene prompt.
Give it facts that nobody else has
Generic input makes generic output. Add three details that come from life: the sound of one particular boiler, the brand of cigarettes, the thing your aunt said at funerals. Specifics crowd out the average.
Ask for the scene without the meaning
"Write what a camera and a microphone would record" removes a surprising amount of explanation in one sentence. Add the inner life back by hand, sparingly.
Ask for the worse version
"Write this exchange as if both people are tired and neither wants to be there" produces better dialogue than any request for realism.
Before and after
Before, the sort of passage chatbots write by default:
"Elara stepped into the old lighthouse, her heart pounding. The air was thick with dust and memory. She let out a breath she didn't know she'd been holding. It wasn't just a building. It was a promise, one her father had never kept. 'You came back,' Elias said softly, his eyes darkening with something she couldn't name."
After, same beat:
"The pump house door had been painted shut. Noreen got it open with the tire iron and stood there, smelling diesel. 'I told them you'd sell,' Wade said from behind the compressor. He had her father's jacket on. 'It's cold,' he said. 'Don't start.'"
Nothing clever happened. The names belong to people of a certain age in a certain place. Nobody explains the jacket. The last line is an argument, not a lesson.
The editing pass no prompt replaces
Set aside twenty minutes per chapter.
- Delete the last sentence of every paragraph and put back only the ones you miss.
- Search for: breath, something, somehow, seemed, felt, a mix of, couldn't help but, and any character's eyes.
- Read the dialogue aloud without the narration. If both people sound alike, change one.
- Break the rhythm. Find the longest run of medium sentences and cut one to four words.
- Change one thing per page to something only you would know.
Will this fool AI detectors?
That is the wrong goal, and no method is reliable. Detectors make mistakes in both directions. Write for the reader, and be honest where a store asks: Amazon KDP counts text as AI-generated "even if you applied substantial edits afterwards" and wants to be told.
What we do about it in BookNova
We have the same enemy at book scale. A chatbot can be corrected in one scene, and then it has thirty more chapters to relapse. BookNova rotates chapter openings and endings so that they do not repeat, can learn your voice from your own pages, and audits the finished draft for twists that leaked too early. It is far from perfect, and you should run the editing pass above on anything it writes. Sample novels are free by email, so judge for yourself, and look at the last third. Prices are here.
Bottom line: AI prose sounds alike because models were rewarded for pleasing strangers. Give it your names, your facts and your bans, put the rules in the settings, and then cut the sentence where it explains itself. That last one does half the work.
FAQ
Why does ChatGPT use so many em dashes?
Older versions learned the habit in training. In our November 2026 test GPT-6 used one or two per 1,000 words. If yours uses more, add a rule to Custom Instructions, as OpenAI suggested.
Why are AI characters always called Elara?
Because names like Elara, Elias and Mara are strongly favored by the data used to align models, according to the Cornell study. In our test three of four models called most of their fantasy heroines Elara. Choose names yourself.
Does Claude sound more human than ChatGPT?
Blind voters prefer its creative writing, and one phrase-count measure found that it uses more stock phrases than GPT-6 Sol. See Claude for Fiction Writing.
Can I train ChatGPT on my style?
You cannot train it, but a description of your voice and samples of your prose in a Project get you much of the way.
Sources, all read November 22, 2026
- BookNova test of four models, November 22, 2026 (160 premises, 12 chapters)
- Hamilton and Mimno: Elias in the Lighthouse, Again?, 2026
- FingerLakes1 on the Cornell study, October 19, 2026
- Namerology: AI has a new crush, October 21, 2026
- Engadget: ChatGPT and em dashes
- Whizi: writing leaderboards, November 2026
- Amazon KDP content guidelines