How we do it
Everyone tells the model to sound human.
We measure whether it did.
A longer prompt is a wish. AIStoryHub runs a closed loop around the model instead: it reads the prose the model just wrote, finds the fingerprints that give machine writing away, and rewrites them from scratch, on every chapter you generate.
The real problem
An AI tell has an altitude.
Catching machine prose is not one problem. It is three, stacked, and each one needs a different instrument. Most tools stop at the first and call it done.
Words
The easy layer. "Palpable." "A testament to." "The silence was deafening." We maintain a 758-entry, community-rated corpus of the words and phrases that give AI away, and we match every one.
Sentences
The participial tail that restates the sentence it ends. The "not X but Y" pivot. These are shapes, not words, but a shape is still a pattern you can pin down and hunt.
Structure
The hard layer. A chapter that closes on a rising sentence reversing its opening image, tying a neat thematic bow. No banned word in sight. The tell is the shape, and it only exists across the whole chapter. You cannot grep a shape. This is where we live.
Why “humanize” fails
Every humanizer reupholsters. Ours refuses to.
Hand a machine draft back to a model and say “make this feel authored,” and it rewords the sentences while keeping the exact images underneath. New fabric, same chair.
We proved it to ourselves with a number we call retention: how much of the original’s distinctive imagery survives a rewrite. High retention is the signature of a reupholster, the words moved and the images did not. The enemy was never bad prose. It was inheritance. Any pass that can see what it is replacing will tend to keep it, so we built a pass that cannot.
The metric nobody else watches
Retention
The words moved. The images stayed. That gap is the tell every other tool ships straight to the reader.
The secret sauce
A closed loop the model can’t game.
After every Pro generation, before you ever see the page, the prose runs through four steps. This is the part no prompt can do.
Scan
Pure detectors, no model, sweep the finished text for structural fingerprints. The rule is a threshold: one use of a construction is a writer's voice, three is an engine. We flag the engine and leave the flourish alone.
Critique
A fast model pass confirms each flagged span and names what it is doing for the scene, so the fix keeps the function and drops only the machine-shaped delivery. It also fences off the two or three most human moments so nothing sands them down.
Redraft from intent
Here is the trick. We regenerate the flagged span without ever showing the model the words it is replacing. It gets the intent and nothing else. A model cannot inherit an image it was never shown. You beat inheritance by making it physically impossible.
Keep only if it helped
The rewrite is scored again and kept only if the machine signal actually dropped and nothing else got worse. Trade one tell for another and the gate rejects it. The loop can never make your prose worse. It is not allowed to.
Detected · same opener ×3
She went.
She crossed the roof at a dead run.
She dropped into the dark.
one is voice. a run is an engine.
A tell, caught
The loop, on one real drumbeat.
- 01Scan flags the run: three sentences opening on the same word.
- 02Critique confirms it and names what it paces, a chase at a dead run.
- 03Redraft rebuilds the beat varied, never once shown the original openings.
- 04The gate keeps the rewrite only if the drumbeat actually quieted.
Before the first word
We box out the model’s worst habits before it writes.
The loop catches what slips through. These constraints keep most of it from happening in the first place, baked into the prompt the model never sees you assemble.
Ending architectures
Six ways to close a chapter, rotated by seed so three drafts of one scene never land the same ending. The rising symbolic bookend, the AI default, is banned outright.
Construction budget
Any single framing shape, "the way X," "as if X," "like X might," is capped at two uses per chapter. Reach for it a third time and it gets rewritten a different way.
One image per beat
Never stack metaphors. Write the single strongest image for the moment and stop. The profound line every paragraph is its own kind of tell.
Thirty voice dials
Sentence complexity, metaphor density, emotional directness, and twenty-seven more, tuned into a plain-language voice the model writes toward before it starts.
What it catches
Five fingerprints, verified and enforced.
Not a vibe check. Five specific structural shapes, each one detected, confirmed, rewritten, and gated so it can only ever get quieter.
Framing engines
The same syntactic frame worn as a crutch, "the way X" four times in two thousand words.
Symbolic bookend endings
The long final sentence that reverses the opening image to tie the scene into a theme.
Detachable epigrams
The quotable standalone aphorism that generalizes past the scene and sounds wiser than the character.
Inherited fossils
An image that survived every prior draft untouched, kept by inertia rather than chosen for this scene.
The anaphora drumbeat
Sentence after sentence opening the same way. "She went. She crossed. She dropped." One is voice. A run is an engine.
And the list keeps growing. Every shape we can name, we can measure. Every shape we can measure, we can cut.
Revision, too
Even your edits lose their inheritance.
Revision is where reupholstering hides best. Reach back through a chapter’s last three drafts and some images are always there, carried untouched through every edit. We find those fossils, ban them by name, and refuse to keep a revision unless it cut at least one. Other tools measure the problem. Ours puts a door on it, and the door only opens if the number went down.
Free model. Bring your own key. Your words stay yours.
Write prose that doesn’t have an accent.
Start on our free model or plug in your own. The whole system runs on every chapter you generate, no setting to flip.