Why Slopsleuth
So the next great voice doesn't get killed by a statistical false positive.
The problem
GPTZero. Originality.ai. Copyleaks. Turnitin. They were built for high-volume screening of college essays and corporate text — and they're useful there. But they apply broad statistical perplexity models to every kind of writing they're given. Fiction is the exception, not the average.
"It was a good fight. Not bad. Just enough. He had not expected the boy to fight at all."
A perplexity-based detector sees that and flags it as AI. The model can't tell the difference between a Nobel Prize–winning declarative voice and a chatbot's hedge fragment. To the algorithm, both are just low-entropy text.
The same fate awaits Hemingway, Cormac McCarthy, Joan Didion, Raymond Carver, Don DeLillo — any author whose voice doesn't fit the statistical average of the training corpus. Minimalism, fragmentation, controlled repetition: all read as machine-like to a generic detector.
A debut novelist gets one shot per submission. A false positive on an agent's screening tool is the end of that submission — sometimes the end of the book's chance entirely. The next Hemingway, writing in 2026 in a small town on a borrowed laptop, gets flagged 73% AI by an algorithm trained on Reddit comments.
When a black-box detector spits out a percentage with no explanation, an agent has no way to defend a rejection — or, worse, an acceptance. So they stop using the tools, and the slush pile becomes a guessing game again.
In the U.S., AI-generated text cannot be copyrighted (USCO, 2023). A book that turns out to be partially AI-written is a book whose protection is partial, whose competitors can reprint freely, and whose royalties may have to be refunded. Every agent quietly fears this.
Tested, not promised
We ran Hemingway's The Sun Also Rises (1926) through Slopsleuth — a novel that perplexity-based detectors routinely misclassify.
Slopsleuth scored it 7 / 100 — LIGHT SIGNALS, with the recommendation "Read with normal attention." Not flagged as AI.
Our Stock AI-Phrasings audit, the strongest discriminator, returned 0.25 per 1000 words — clean. By contrast, a Claude-generated thriller written cold scored 23 / 100 with 4.28 stock-phrasings per 1000 words — LOAD.
The gap is real and reproducible. Try the samples in the app yourself.
We don't punish minimalist prose. We don't punish stylistic risk. We hunt the specific tics that current language models leave behind — every flag is shown to you with the actual passage.
I championed this book, sold it to a publisher, and now we're learning it was AI-generated. That phone call is bad. The lawsuit that follows is worse. Slopsleuth gives you a defensible, transparent read — evidence-based, not vibes-based.
The art of writing fiction is being commodified at industrial scale. Our work is to draw the line between a writer who used a model as a thesaurus and a writer who pasted a chapter from one. We err on the side of the human.
Click any sample manuscript. See the audit. Make up your own mind.
Launch Slopsleuth →