AI Humanizer for Fiction and Ghostwriters: StealthGPT vs Pangram
StealthGPT Super took our AI-written literary-fiction and memoir samples from 100% AI in Pangram 4.1 to 0% AI and “Human Written.” A romance scene moved from 100% AI to 5% AI and 95% human, with a “Mostly Human” label. These are fresh November 15, 2026 results from the first rewrite of each passage, with the original text, output and detector reports saved.
For fiction writers and ghostwriters, the useful demonstration is the writing itself. Below, you can see what happened to a short literary scene, a second-chance romance and a fictional memoir anecdote. The examples show substantial rewrites across distinct voices, with a result table that keeps all three outcomes visible.
Want to try the same writing workflow? Explore the StealthGPT AI Humanizer, or open the humanizer with a draft you want to revise.
The Pangram results: three writer-focused examples
These are three fiction and ghostwriting passages from our recorded genre run. Every original was written specifically for the test by an AI research agent. We did not copy a published novelist’s scene or run a passage from any writer mentioned in the news.
| Writing sample | Original AI classification | After one Super rewrite | Pangram result |
|---|---|---|---|
| Literary fiction | 100% AI | 0% AI · 100% human | Human Written |
| Romance fiction | 100% AI | 5% AI · 95% human | Mostly Human |
| Fictional ghostwritten memoir | 100% AI | 0% AI · 100% human | Human Written |
Two of these three outputs met our strict pass definition: no text classified as AI, no text classified as AI-assisted, and a Human Written result. The romance output improved substantially but did not meet that definition. The percentages describe Pangram’s classification of the scanned text; they are not probabilities that a real person wrote it.
Our full StealthGPT-versus-Pangram benchmark keeps the full recorded test material together. The writer-focused examples here are the three predeclared fiction and memoir fixtures within that broader test.
Why writers are paying attention to Pangram
Download the original passages, first Super outputs and saved Pangram reports to inspect the complete test record.
AI detection has moved beyond classroom essays into the business of publishing. An October 31 NPR report described a nonfiction book-proposal sample being screened and the pressure this creates for writers and editors. The practical concern is familiar to a working ghostwriter: a client may evaluate the delivered prose through a detector as well as by reading it.
Fiction has been part of the discussion too, but the latest reporting deserves to be described accurately. On November 12, The Guardian reported that the Caine Prize had cleared a winning story after an investigation that considered drafts, metadata and interviews. The prior accusation had relied on Pangram. That was a cleared writer, not a newly confirmed AI catch.
That coverage makes original, inspectable writing examples especially useful. We created our own passages, humanized them once, and checked the exact outputs in the live Pangram 4.1 dashboard. Writers can examine those saved texts and reports alongside the news context.
Literary fiction: from 100% AI to Human Written
Our literary fixture follows Mara as she returns to her father’s closed bakery in Bellhaven. Its details are deliberately specific: a stopped clock, a blue tin of family photographs, a childhood stool and a folded railway timetable. This is a scene built around memory and an unresolved decision, rather than an informational paragraph dressed up with adjectives.
Original opening:
By the time Mara returned to Bellhaven, the clock above her father's bakery had stopped at eleven seventeen. It had always run seven minutes fast, a small private disagreement with the rest of the town.
First Super rewrite:
The photograph was of her mother coming through the door with a tray of rolls, perhaps they were going out to be sold in the courtyard, and Mara remembered the chipped cup on the wall behind her, remembered the dress, but not who had taken the photograph.
Super moves the opening into the photograph and Mara’s recollection. The bakery, character and central objects remain recognizable, but the narrative entry point changes. That is the distinction a fiction writer can assess directly: this is a structural rewriting option, not simply a row of substituted words.
Pangram 4.1 result: the original was 100% AI; the first rewrite was 0% AI, 0% AI-assisted and 100% human, labeled Human Written. The saved report is the evidence for that result, while the complete passage lets you judge the writing itself.
Pangram 4.1 · saved first Super output. Open the full-size scan screenshot.
Exact test drafts: Read the original AI draft · Read the first StealthGPT Super output.
Romance fiction: a large improvement with a result worth showing
The romance fixture pairs Nina and Theo Mercer at a winter fundraiser in Willow House. They have a four-year history, a leaking roof to raise money for and sixty guests to feed. The scene ends with a dinner invitation and a table-for-two callback. Its challenge is different from the literary sample: dialogue and emotional timing carry the story.
Original ending:
When Theo asked whether she would have dinner with him after the fundraiser, Nina smiled at the long rows of empty chairs. This time, she said, he would have to reserve a table for two.
First Super rewrite:
‘Will you come to dinner with me after the fundraiser?' Theo asked. Nina looked around the hall at all the empty chairs. ‘If you want to ask me out for dinner, Theo Mercer, you're going to have to book a table for two.'
The rewrite turns the indirect invitation into a spoken exchange while keeping the dinner setup and callback. It gives the writer an alternative way to stage the same romantic moment. Read the complete source and rewrite together to decide which lines fit the characters you are developing.
Pangram 4.1 result: the original was 100% AI. The first rewrite was 5% AI and 95% human, labeled Mostly Human. It remains in the table as the first output, clearly distinguished from the two fully clean classifications.
Pangram 4.1 · saved first Super output. Open the full-size scan screenshot.
Exact test drafts: Read the original AI draft · Read the first StealthGPT Super output.
Ghostwritten memoir: a more conversational anecdote, then a clean scan
The memoir fixture uses a fictional furniture-maker’s first delivery van and an unfinished cabinet to tell a story about responsibility. The narrator, daughter Elsie and workshop in North Quay are invented for the demonstration. There is no real client’s life story, interview transcript or private material behind the passage.
Original opening:
I was twenty-three when I bought my first delivery van, a white vehicle with a dented rear door and a heater that worked only when the road turned downhill. My daughter, Elsie, called it the snow truck because paint flaked from the roof whenever we shut the doors.
First Super rewrite:
The first delivery van I ever owned was white. It had a dented back door and the heater only worked going downhill. My daughter, Elsie, nicknamed it the snow truck because every time you closed the doors there would be a flurry of white paint flakes from the roof.
The new opening separates the description into shorter beats and makes the anecdote more conversational. The white van, damaged door, unreliable heater and Elsie’s nickname carry through this excerpt. You can inspect those actual sentences rather than rely on a general claim about “natural voice.”
Pangram 4.1 result: the original was 100% AI; the first Super rewrite returned 0% AI, 0% AI-assisted and 100% human, labeled Human Written.
Pangram 4.1 · saved first Super output. Open the full-size scan screenshot.
Exact test drafts: Read the original AI draft · Read the first StealthGPT Super output.
How to use an AI humanizer in a real writing workflow
The writer owns the scene, the facts and the final voice. Our saved outputs make that role visible: the literary rewrite changes its ending, and the memoir rewrite adds a customer-follow-up episode. Review those creative changes deliberately; for real ghostwriting, confirm new incidents against the subject’s approved notes.
Start with the writing goal. Bring an AI-assisted scene, sample chapter or anecdote and identify what it needs: a different rhythm, more conversational delivery or another narrative approach.
Keep the original beside the rewrite. Names, relationships, chronology and key events deserve a deliberate check. In ghostwriting, compare factual statements with the approved notes or transcript.
Generate a useful alternative. We used Super once for each test passage. Read its version as an editing option and decide which language belongs in the project.
Check the exact version you are evaluating. Our Pangram reports correspond to the saved, unedited first outputs. If you later revise a passage, that is a new text and deserves its own assessment.
Make the final editorial decisions. Retain the dialogue, images and pacing that work. Repair continuity or remove invented details before a draft becomes a finished piece.
This workflow leaves room for both the product’s strength and the writer’s judgment. StealthGPT supplies another version of the prose. A human writer decides whether the new scene still says what the project needs it to say.
The settings and evidence behind these examples
Test date: November 15, 2026.
Humanizer: StealthGPT’s web AI Humanizer, Super model; Quality Checks on, Em Dash Remover on, Formatter off.
Attempts: one Generate click and the first returned output for each predeclared sample. No retries or manual edits before these scans.
Detector: the live Pangram dashboard; the saved reports identify Pangram 4.1.
Materials: original AI-written English fixtures, with source text, first output and before/after reports retained. These are short passages, not whole-book tests.
Reporting: all three writer-focused outcomes are shown, including the romance result that missed the strict clean-scan definition. The full scan record lives in the benchmark.
For the broader product comparison, see StealthGPT’s Pangram workflow and results. For the complete texts and the other tested writing genres, use the Pangram benchmark evidence hub.
Questions fiction writers and ghostwriters ask
Can Pangram detect AI-written fiction?
It detected all three original AI-written writer fixtures in this test as 100% AI. After one StealthGPT Super rewrite, two received Human Written results and the third received Mostly Human with 5% AI. The useful answer is the recorded before-and-after evidence, not a promise about every story.
Which StealthGPT model did you use for these writing examples?
We used Super in the web AI Humanizer with Quality Checks on. Each passage was generated once and the first returned output was scanned in Pangram. The screenshots and saved files correspond to those first outputs.
Does a Human Written result change where the text came from?
No. These passages began as original AI-written test fixtures and were rewritten with StealthGPT. Human Written is Pangram’s classification of the scanned output. We keep the text’s origin explicit so readers can understand exactly what was tested.
Can I use this process for a novel or a ghostwriting project?
You can use a humanizer to develop alternative prose for a scene or section, then review voice, continuity and facts within your normal editing process. This test covers short passages. It does not report a full novel, a complete memoir or a publisher’s submission decision.
Try it on the kind of writing you actually do
One Super rewrite took our literary-fiction and fictional-memoir passages from 100% AI to Human Written in Pangram 4.1. The romance example reached 95% human. The original prose, first rewrites and saved reports give you a concrete way to evaluate this workflow for your own writing.
Bring your next draft to the StealthGPT AI Humanizer, or explore the humanizer’s writing features before you begin.