Finding the AI in writing, and helping people sound like themselves
Most people now write with AI. The result often sounds the same: the same openings, the same rhythm, the same tidy wrap-ups. Readers notice, and they trust the writing less. Desloppy reads a piece of writing, scores how much of it reads like AI, and points to the exact phrases that give it away, with a plain fix for each. It combines a detection model we train ourselves with a growing library of the habits AI writing shows. The model keeps learning from the newest AIs, and a new version only replaces the current one when it does better on a test it has never seen. On real everyday AI posts it caught every one in our test, and it rarely mistakes human writing for AI.
1The problem
AI tools write fast and well, but they write alike. A promotion post, a cover letter and a sales email all start to share the same words and the same shape. People who use AI to help them write still want the result to sound like them. People who read it want to know a person stands behind it.
Most AI detectors give one number and stop there. A number alone doesn't help a writer. Desloppy is built to answer the next question too: which lines, and what do I change?
2Our approach
Desloppy looks at writing in two ways and puts the results together.
A detection model. A language model trained to tell human writing from AI writing. It scores the whole text and each part of it, so we can show which passages read most like AI.
A library of AI habits. Phrases, structures and patterns that AI writing leans on, gathered from studying large amounts of AI and human writing. Each habit comes with a plain fix, so the advice is specific: not "sounds robotic", but which words to change and how.
Both are tuned to the kind of writing: an email, a social post, a cover letter, a slide deck, an essay or a story each has its own normal style, and what counts as a warning sign in one can be fine in another.
3Always learning
New AI models come out every few weeks, and each one writes a little differently. A detector trained once goes stale. So Desloppy runs as a loop that keeps going on its own:
This means Desloppy gets better without anyone having to remember to update it, and a bad update can never reach you. Your writing is never used for training.
4Results
These come from our own model's release test in October 2026. The test set is held back and never used for training.
| Test | Result |
|---|---|
| Everyday LinkedIn posts written by five AIs (Claude, ChatGPT, Gemini, Qwen, GPT-OSS) | 101 of 101 caught |
| AI text that was paraphrased, rewritten or run through a "humanizer" | 97% caught |
| AI from model families it was never trained on | 91% caught |
| Hard human writing (student essays, short texts, AI-polished human text) wrongly called "likely AI" | 0.3% |
5What it can't do
- A score is a probability, not proof. Never use Desloppy alone to accuse anyone of using AI.
- Some writing is hard to judge. Very short texts, lists and heavily formatted text give the model little to go on, and Desloppy says so.
- Long fiction carefully styled by an author with AI is still hard for every detector, ours included. It is a focus of our weekly training.
- English only, for now. Text in other languages gets no verdict and uses none of your words.
6Privacy
Your text is checked in memory and dropped as soon as the result is sent. We don't store it and we never train on it. We keep what an account needs, like your email, scrambled copies of your keys and how many words you have used. The privacy policy lists all of it.
7Get in touch
Questions, ideas or a text we got wrong? Write to support@desloppy.com. Wrong calls help us most.