facts, not marketing
How Neuroslop finds AI slop
In short
Neuroslop is a deterministic detector of careless machine text plus a humanizer. It scores liveliness from 0 to 100 and breaks the text down by concrete markers. Checking is free and unlimited, and we don't store your text. This is not an “AI police”: well-written text passes, even if a model wrote it. What gets flagged is sloppy work with the usual traces.
What the detector looks for
- AI clichés and crutch connectors. Hundreds of rules across 29 categories: “in today's world”, “plays a key role”, “opens new horizons”, bureaucratese, empty intensifiers.
- Windowed repetition. Word repeats are counted in a sliding window of about 45 words, not across the whole text. Topical words and proper nouns are excluded, so an article about a single company doesn't get false penalties.
- Monotone rhythm. Several sentences of identical length in a row — a typical trace of generation.
- Structural signatures. Antitheses, three-item lists, machine punctuation like em dashes in every paragraph.
Why there are few false positives
- Register before penalties. The detector first reads the register of the text. Dry journalism and business prose aren't treated as slop.
- Genre awareness. News, academic writing, a contract, marketing and a personal post are judged by different yardsticks.
- Corpus safeguard. The dictionaries are checked against a corpus of living literary text, so classic turns of phrase don't become false markers.
- Weak single markers. Clusters of findings decide the verdict. One word never sinks a text.
Self-learning
The cliché dictionaries grow automatically: user complaints, distillation from an LLM, and external sources. A new phrase enters the rules only after confirmation from several different users — a defense against gaming. Phrases are filed into the dictionary of their own language.
Humanization
The model rewrites the text on our prompt: it removes the flagged clichés, breaks the monotone rhythm, and keeps the meaning, genre and the author's voice. Humanization is not neutralization: emotional text stays emotional. The before/after score shows instantly. Checking is free; humanization gives 3 free rewrites a day, then credit packs — Standard and Premium.
Languages
The interface is translated into 20 languages, and both the breakdown and the rewrite work in each. The detector identifies the text’s language by script and stop-words and applies that language’s dictionary. Text that mixes several languages is handled too.
Where it works
- The website: free checking and a workspace with humanization.
- The Telegram bot @AITextChecker_bot: the same core and plans, paid with Telegram Stars.
- A public HTTP API for developers: check and humanize by key.