How to determine who wrote the text
Orhuman detector methodology. We don't just look for "AI probability", we determine the nature of authorship.
Authorial
One distinct voice, uneven rhythm, personal markers, and details.
Editorial
Dry, professional text. Requires confirmation from online sources.
Generated
Machine clichés, "water", structural symmetry, and sterility.
Hybrid
Machine framework with traces of manual editing or splices (seams).
Why we look for the original source (Fact-checking)
Classic AI detectors often make mistakes: if you paste a Wikipedia article or a legal text into them, they will show 99% AI. Why? Because encyclopedias and neural networks write in the same dry, faceless, and structured manner.
To avoid confusing a formal document with a generated reference, the algorithm first checks for matches in open sources. The "Editorial Text" status is assigned only with a confirmed source. If the text has an encyclopedic style but is not found online, the decision is made based on a combination of structural features and the density of machine patterns.
Why is a proofread text sometimes identified as AI?
Neural networks were trained on perfectly polished, edited articles. Therefore, when a good commercial author runs a text through a typographer, cleans up the syntax, and maintains a flat structure — old algorithms looked at this "perfection" and screamed: "It's a machine!". By bringing the text to a corporate ideal, a person erases their biological traces.
What we DO NOT look at:
Perfect punctuation, typographic quotes, and smooth syntax are not determining factors. In the new core version, these are merely secondary signs of good proofreading, not a verdict. They affect the score only in combination with algorithmic patterns and semantic hallucinations.
What gives away the neural network:
We look for semantic hallucinations: the fear of taking a hard stance, the desire to please everyone, the over-explanation of the obvious, and template "red flags" in the structure of the argumentation.
Pattern Library
The algorithm looks for generation anomalies the way a professional editor does. Markers are divided into logical groups, from the most critical (Red Flags) to auxiliary ones. Select a category for a quick jump:
Editing & Genre
Signs of professional editing and genre. Dryness, density, and depersonalization are normal for normative and reference texts. Not a sign of machine generation by themselves.
Absence of both human "noise" and generative "water". Maximum dry facts per sentence.
Intentional avoidance of "I/we" pronouns, use of passive voice and detached presentation.
Use of strict phrasing without attempting to over-explain them for casual readers (unlike AI).
Translation
Signs of translation or calque from a foreign language. Considered separately from machine generation.
Transliteration or duplication of foreign names, brands, and terms in Latin characters.
Using English syntactic constructions and heavy piling of nouns in the genitive case.
Using full, cumbersome names of foreign professions instead of familiar abbreviations.
Using foreign measurement systems (miles, pounds, fahrenheit) without converting to common formats.
Quotes lack natural conversational elements and sound like a verified written press release.
Classic dry structure of Western agencies: «fact → historical context → expert opinion».
Human Author
Markers of individual authorship. Traces of real thought processes, associations, and temporal anchors.
The text shows traces of how the author reaches a conclusion or corrects an opinion right while writing.
Temporary departure from the main topic due to a sudden thought or memory, typical of living thinking.
The thought is cut off in the middle of a logical chain, as the conclusion is already obvious to the author.
Lack of explanations for specific names or terms that are perceived by the author as basic.
A recurring atypical systemic error or filler word peculiar to a specific person.
Mentioning a random object or fact that carries no practical or emotional benefit for the article.
Using the "here and now" context (anchoring to the current moment of writing, not an archival date).
Broken case agreements or duplicate words left after changing a sentence on the fly.
Intonation and hidden assumptions show that the author is addressing a narrow, understanding circle of people.
Red Flags
Critical patterns of automatic generation. Stable templates and structural markers.
Accidental inclusion of an English word in a translated text without an apparent stylistic need.
Repeated use of the algorithm's base construct: "It's not just [A], it's [B]".
Starting the text with a direct announcement of the action plan ("In this article, we will consider...").
Ending a technical or personal text with a template question to the audience, typical of SMM scripts.
Thinking
Behavioral features of generative models: symmetry of argumentation and neutralization of positions.
The main conclusion is given in the first paragraph, and the rest of the text only formally justifies it.
The text tries to accommodate all points of view, avoiding taking a specific, radical, or controversial position.
Using exact dates, full names, and numbers in an informal context where a human would write approximately.
Every word works strictly to reveal the topic. There are no accidental side details typical of human memory.
Detailed decoding of simple terms and abbreviations, even if the text is intended for professionals.
Artificial balancing of pros and cons with proportional blocks of text.
Emotions & Tone
Analysis of emotional dynamics and changes in intonation.
Feelings are named by words ("this caused panic") but are not reflected in the rhythm of the sentences.
The style does not change from beginning to end. No signs of authorial acceleration or fatigue.
The narrative is aimed at an abstract average reader, ignoring professional slang or shared pain.
Immediately after a joke or complex metaphor comes its logical decoding, neutralizing the comic effect.
Hybrid: Injection
Signs of a hybrid text: local stylistic and structural gaps.
One paragraph or block differs sharply from the rest by the presence of living experience or conversational syntax.
Sudden transition from cold narrative to vivid emotion (pain, rage) and an immediate return back.
Sharp change in average length and structure of sentences in a short isolated section of text.
Words of human uncertainty ("probably", "around") inserted into an array of absolutely accurate encyclopedia data.
If a suspicious paragraph is removed, the surrounding text closes perfectly without losing the logic.
Falsification
Attempts to imitate human style using intentional stylistic and orthographic insertions.
Excessive and unnatural use of conversational slang in every sentence (hyperstylization).
Rough conversational vocabulary embedded into a complex, strictly academic sentence framework.
A single, mathematically calibrated spelling error embedded in a surgically flawless text.
Introductory words of uncertainty ("perhaps", "probably") dissonant with a rigid and categorical conclusion.
A "real life" story that proves the theoretical thesis of the article too straightforwardly and like a textbook.
Key Questions
Control questions to assess the presence of a subject of authorship.
The material lacks professional or personal biography. It could have been published by anyone.
Absence of specific, hard-to-explain memories that usually accompany personal experience.
The text contains no logical gaps: the author does not miss a single detail they might consider obvious.
Language
Lexical and syntactic features. Auxiliary level of analysis.
Absence of minor typos, broken cases, or accidental repetitions typical of fast manual typing.
Complex, grammatically verified constructions without stylistic roughness typical of conversational speech.
Sentences have approximately the same length and structure, creating a machine-like rhythm.
Using broad, voluminous phrases ("key aspect", "opens new horizons") that carry no concrete value.
Using the most obvious and clichéd comparisons based on the frequency statistical analysis of the vocabulary.
Structure
Structural organization and visual symmetry of paragraphs. Secondary feature.
Text is divided into visually equal blocks. Mathematical evenness is often a sign of generation.
The last paragraph paraphrases the first. Algorithms often use this template for formal summaries.
Paragraphs are not connected by smooth logical transitions and can be rearranged without losing meaning.
Using evasive phrases ("sometimes", "in some cases") to avoid harsh, unambiguous judgments.
The text ends with a mandatory inspiring moral or hope for the best, regardless of context.
Typography
Typographic and technical artifacts. Used as auxiliary signals.
Using the correct long dash (U+2014) instead of a standard keyboard hyphen.
Using a single U+2026 character instead of three consecutive dots typed manually.
Automatic opening and closing of correct typographic quotes instead of straight typewriter quotes.
Using hidden anchor characters before percent signs, currencies, or initials.
Perfect formatting of punctuation marks over large volumes of text, without accidental double spaces.