How This Client-Side Stylometric AI Detector Works
Every calculation in this tool runs directly inside your web browser. Your text is never transmitted to an external server, saved in a database, or used to train third-party language models. When you click Analyze Text, our local stylometric engine segments the draft and evaluates four complementary linguistic dimensions using a non-linear asymptotic scoring model:
- Sentence Rhythm and Dynamic Burstiness: The analyzer evaluates how naturally sentence lengths fluctuate across the passage, measuring dispersion and rhythm shifts between adjacent clauses rather than relying on simple word-count averages.
- Discourse Clichés and Excess Vocabulary Density: The passage is cross-referenced against empirical distributions of stock LLM openings, formulaic transitions, and stylistic vocabulary clusters that appear disproportionately in machine-generated prose across any topic.
- Rhetorical Symmetry and Lexical Variety: Using a moving-window lexical model, the detector measures vocabulary turnover, repetitive clause openings, symmetrical enumerations, and contrastive rhetorical templates common in both corporate outputs and reflective school essays.
- Structural and Compositional Architecture: Beyond surface punctuation and synthetic list formatting, the engine inspects macro-level paragraph balance, heading-to-body framing, and textbook essay composition patterns.
A strict minimum threshold of 60 words is required before generating a diagnostic report. Attempting to determine authorship from one or two isolated sentences is statistical guesswork, so the analyzer waits until enough text is available to measure real dispersion.
Claude Shannon, Zipf’s Law, and the Thermodynamics of Smooth Prose
To understand why machine-generated text feels uncannily flat even when the grammar is spotless, go back to Claude Shannon’s 1948 paper on information entropy. In natural human language, word frequencies roughly follow Zipf’s law, where a handful of function words dominate and a long, erratic tail of rare words carries specific meaning. When a human writes, conditional entropy swings wildly. A writer might construct a long analytical clause, follow it with a sharp fragment for emphasis, and then drop in an unexpected colloquial verb, a specific date, or a concrete observation that has a near-zero prior probability in generic prose.
Autoregressive language models do the opposite. At each step, the decoder predicts a probability distribution over the next token and trims the tail using temperature scaling and nucleus sampling. By design, that sampling filter shaves off high-entropy surprises to prevent incoherence. The side effect is thermodynamic smoothing: sentence lengths regress toward a safe middle band, paragraphs come out looking like bricks from the same mold, and vivid verbs get replaced by the same polished academic vocabulary over and over. Even when a chatbot writes a continuous philosophical or school essay without bullet points, its internal cadence remains abnormally symmetrical. Respecting Landauer’s limit in physics also means you do not need to burn megawatts of electricity in a liquid-cooled GPU data center just to measure stylometric variance—clean mathematical analysis runs right on your device.
From the 1963 IBM 7090 Federalist Papers Study to Modern Excess Vocabulary
Counting word frequencies to unmask an anonymous author is older than the microprocessor. Between 1963 and 1964, statisticians Frederick Mosteller and David L. Wallace fed the disputed Federalist Papers into an IBM 7090 mainframe. Alexander Hamilton and James Madison wrote with nearly identical eighteenth-century sentence lengths, so average clause size was useless. Mosteller and Wallace solved the puzzle by ignoring grand political nouns and looking at humble function words and habitual discourse markers that an author uses unconsciously.
Sixty years later, the exact same statistical leak exposes modern chatbots. In a 2025 study published in Science Advances, Dmitry Kobak and colleagues analyzed more than 15 million biomedical abstracts indexed on PubMed to see how scholarly writing changed after late 2022. They demonstrated that specific ornamental verbs, adjectives, and discourse frames suddenly spiked across hundreds of thousands of papers. A single elevated verb or formal transition means nothing on its own, because humans use formal language too; however, when lexical excess clusters together with symmetrical triads, contrastive negations, and uniform paragraph framing, the cumulative probability curve rises sharply.
The Black-Box Oracle Cult: Why Stylometry Is a Thermometer, Not a Court Verdict
Somewhere around 2023, a strange superstition took hold in classrooms and editorial offices: the belief that a web form can look at a short essay, flash a red percentage on screen, and serve as an infallible lie detector. That is nonsense. Even more entertaining are commercial portals that demand your email address, ask for a credit card, and upload your confidential manuscript to their cloud servers just to hand you a single mysterious number with zero explanation of what triggered it.
Stylometry measures surface texture, not human conscience or legal guilt. A rigid corporate press release, an encyclopedic summary, a legal contract, or a student who strictly follows a five-paragraph textbook template will naturally register higher structural uniformity. Conversely, an AI draft that a human editor cuts up, enriches with concrete facts, and strips of stock clichés will score as human—because at that point, the prose rhythm genuinely is human. Use the four diagnostic bars above to spot robotic monotony and overused rhetorical habits in your own writing, never as sole evidence to accuse someone without proof.